diff --git "a/MI300x_llama8b_pp1_tp1/profiling_bs1_pl512.json" "b/MI300x_llama8b_pp1_tp1/profiling_bs1_pl512.json" --- "a/MI300x_llama8b_pp1_tp1/profiling_bs1_pl512.json" +++ "b/MI300x_llama8b_pp1_tp1/profiling_bs1_pl512.json" @@ -130,24 +130,24 @@ { "entry": { "name": "LlamaForCausalLM", - "cuda_time_us": 26679.669000000005, - "pct_cuda_time": 98.21927135343087, + "cuda_time_us": 27263.346, + "pct_cuda_time": 98.26510213414547, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", - "cuda_time_us": 43.384, - "pct_cuda_time": 0.1597150574992982, + "cuda_time_us": 44.466, + "pct_cuda_time": 0.1602685169860263, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectLargeIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, unsigned int, long)", - "cuda_time_us": 43.384, - "pct_cuda_time": 0.1597150574992982, + "cuda_time_us": 44.466, + "pct_cuda_time": 0.1602685169860263, "invocations": 1 }, "children": [] @@ -157,24 +157,24 @@ { "entry": { "name": "LlamaDecoderLayer", - "cuda_time_us": 26622.225, - "pct_cuda_time": 98.00779542306505, + "cuda_time_us": 27204.659, + "pct_cuda_time": 98.05357695858753, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cuda_time_us": 930.4409999999999, - "pct_cuda_time": 3.4253512312074617, + "cuda_time_us": 936.4179999999998, + "pct_cuda_time": 3.3751253573296616, "invocations": 64 }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, unsigned long, unsigned long)", - "cuda_time_us": 6.449, - "pct_cuda_time": 0.023741526964156694, + "cuda_time_us": 6.929, + "pct_cuda_time": 0.02497415000666073, "invocations": 1 }, "children": [] @@ -182,8 +182,8 @@ { "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", - "cuda_time_us": 923.9919999999998, - "pct_cuda_time": 3.4016097042433047, + "cuda_time_us": 929.4889999999999, + "pct_cuda_time": 3.350151207323001, "invocations": 63 }, "children": [] @@ -193,24 +193,24 @@ { "entry": { "name": "LlamaAttention", - "cuda_time_us": 9994.584, - "pct_cuda_time": 36.79433796426254, + "cuda_time_us": 10584.966, + "pct_cuda_time": 38.151324678799774, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cuda_time_us": 2566.761, - "pct_cuda_time": 9.449344935966167, + "cuda_time_us": 2594.1040000000003, + "pct_cuda_time": 9.34991231474652, "invocations": 32 }, "children": [ { "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", - "cuda_time_us": 2566.761, - "pct_cuda_time": 9.449344935966167, + "cuda_time_us": 2594.1040000000003, + "pct_cuda_time": 9.34991231474652, "invocations": 32 }, "children": [] @@ -220,16 +220,16 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cuda_time_us": 1200.7499999999998, - "pct_cuda_time": 4.420474259917995, + "cuda_time_us": 1202.9600000000003, + "pct_cuda_time": 4.335820968684168, "invocations": 32 }, "children": [ { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 846.691, - "pct_cuda_time": 3.1170316648796397, + "cuda_time_us": 844.4140000000004, + "pct_cuda_time": 3.043515933572583, "invocations": 128 }, "children": [] @@ -237,8 +237,8 @@ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", - "cuda_time_us": 354.05899999999997, - "pct_cuda_time": 1.3034425950383555, + "cuda_time_us": 358.54600000000005, + "pct_cuda_time": 1.2923050351115861, "invocations": 32 }, "children": [] @@ -248,16 +248,16 @@ { "entry": { "name": "Attention", - "cuda_time_us": 4637.619, - "pct_cuda_time": 17.07305885222289, + "cuda_time_us": 4642.354, + "pct_cuda_time": 16.732406578152904, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", - "cuda_time_us": 312.95699999999994, - "pct_cuda_time": 1.1521285554538045, + "cuda_time_us": 314.276, + "pct_cuda_time": 1.1327429596613232, "invocations": 32 }, "children": [] @@ -265,8 +265,8 @@ { "entry": { "name": "attn_fwd", - "cuda_time_us": 4324.662, - "pct_cuda_time": 15.920930296769086, + "cuda_time_us": 4328.078, + "pct_cuda_time": 15.599663618491583, "invocations": 32 }, "children": [] @@ -276,16 +276,16 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cuda_time_us": 1589.4540000000006, - "pct_cuda_time": 5.851459916155487, + "cuda_time_us": 2145.5480000000002, + "pct_cuda_time": 7.733184817216182, "invocations": 32 }, "children": [ { "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", - "cuda_time_us": 1589.4540000000006, - "pct_cuda_time": 5.851459916155487, + "cuda_time_us": 2145.5480000000002, + "pct_cuda_time": 7.733184817216182, "invocations": 32 }, "children": [] @@ -297,24 +297,24 @@ { "entry": { "name": "LlamaMLP", - "cuda_time_us": 15697.2, - "pct_cuda_time": 57.78810622759506, + "cuda_time_us": 15683.275000000005, + "pct_cuda_time": 56.5271269224581, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cuda_time_us": 9206.382000000001, - "pct_cuda_time": 33.892629321650936, + "cuda_time_us": 9180.412, + "pct_cuda_time": 33.08889975623442, "invocations": 32 }, "children": [ { "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", - "cuda_time_us": 9206.382000000001, - "pct_cuda_time": 33.892629321650936, + "cuda_time_us": 9180.412, + "pct_cuda_time": 33.08889975623442, "invocations": 32 }, "children": [] @@ -324,16 +324,16 @@ { "entry": { "name": "SiluAndMul", - "cuda_time_us": 838.0, - "pct_cuda_time": 3.0850363771070417, + "cuda_time_us": 829.6049999999998, + "pct_cuda_time": 2.9901399503933863, "invocations": 32 }, "children": [ { "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", - "cuda_time_us": 838.0, - "pct_cuda_time": 3.0850363771070417, + "cuda_time_us": 829.6049999999998, + "pct_cuda_time": 2.9901399503933863, "invocations": 32 }, "children": [] @@ -343,16 +343,16 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cuda_time_us": 5652.818, - "pct_cuda_time": 20.81044052883708, + "cuda_time_us": 5673.258, + "pct_cuda_time": 20.448087215830284, "invocations": 32 }, "children": [ { "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", - "cuda_time_us": 5652.818, - "pct_cuda_time": 20.81044052883708, + "cuda_time_us": 5673.258, + "pct_cuda_time": 20.448087215830284, "invocations": 32 }, "children": [] @@ -366,16 +366,16 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cuda_time_us": 14.06, - "pct_cuda_time": 0.05176087286649762, + "cuda_time_us": 14.221, + "pct_cuda_time": 0.051256658571903924, "invocations": 1 }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", - "cuda_time_us": 14.06, - "pct_cuda_time": 0.05176087286649762, + "cuda_time_us": 14.221, + "pct_cuda_time": 0.051256658571903924, "invocations": 1 }, "children": [] @@ -387,16 +387,16 @@ { "entry": { "name": "LogitsProcessor", - "cuda_time_us": 239.595, - "pct_cuda_time": 0.8820516596336058, + "cuda_time_us": 238.554, + "pct_cuda_time": 0.8598186434823126, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", - "cuda_time_us": 4.806, - "pct_cuda_time": 0.017692941322644918, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.01660137609044297, "invocations": 1 }, "children": [] @@ -404,8 +404,8 @@ { "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", - "cuda_time_us": 234.789, - "pct_cuda_time": 0.8643587183109609, + "cuda_time_us": 233.948, + "pct_cuda_time": 0.8432172673918696, "invocations": 1 }, "children": [] @@ -415,16 +415,16 @@ { "entry": { "name": "Sampler", - "cuda_time_us": 244.111, - "pct_cuda_time": 0.8986769869355334, + "cuda_time_us": 242.78799999999998, + "pct_cuda_time": 0.8750792223722248, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy DtoD (Device -> Device)", - "cuda_time_us": 35.246, - "pct_cuda_time": 0.12975559922137805, + "cuda_time_us": 33.604, + "pct_cuda_time": 0.12111868044794737, "invocations": 7 }, "children": [] @@ -432,8 +432,8 @@ { "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 2.643, - "pct_cuda_time": 0.009730013299157412, + "cuda_time_us": 2.763, + "pct_cuda_time": 0.009958663078135893, "invocations": 1 }, "children": [] @@ -441,8 +441,8 @@ { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2})", - "cuda_time_us": 13.9, - "pct_cuda_time": 0.051171844441274324, + "cuda_time_us": 13.218, + "pct_cuda_time": 0.047641552141440544, "invocations": 1 }, "children": [] @@ -450,8 +450,8 @@ { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", - "cuda_time_us": 78.236, - "pct_cuda_time": 0.28802017422356385, + "cuda_time_us": 78.757, + "pct_cuda_time": 0.28386334710269584, "invocations": 1 }, "children": [] @@ -459,8 +459,8 @@ { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", - "cuda_time_us": 44.786, - "pct_cuda_time": 0.1648764190753174, + "cuda_time_us": 44.746, + "pct_cuda_time": 0.1612777191799742, "invocations": 1 }, "children": [] @@ -468,8 +468,8 @@ { "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 5.006, - "pct_cuda_time": 0.018429226854174047, + "cuda_time_us": 4.926, + "pct_cuda_time": 0.017754750026383426, "invocations": 1 }, "children": [] @@ -477,8 +477,8 @@ { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 6.89, - "pct_cuda_time": 0.02536503656117842, + "cuda_time_us": 7.17, + "pct_cuda_time": 0.02584278475216589, "invocations": 1 }, "children": [] @@ -486,8 +486,8 @@ { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4, 4> >(at::native::ReduceOp, unsigned int, long, 4, 4>)", - "cuda_time_us": 52.157, - "pct_cuda_time": 0.19201222233982335, + "cuda_time_us": 52.718, + "pct_cuda_time": 0.19001114735909086, "invocations": 1 }, "children": [] @@ -495,8 +495,8 @@ { "entry": { "name": "Memcpy DtoH (Device -> Host)", - "cuda_time_us": 5.247, - "pct_cuda_time": 0.019316450919666644, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.01761057828439087, "invocations": 1 }, "children": [] @@ -508,18 +508,18 @@ { "entry": { "name": "LlamaForCausalLM", - "cpu_time_us": 34403.25, - "cuda_time_us": 26679.669000000005, - "pct_cuda_time": 98.21927135343087, + "cpu_time_us": 35771.095, + "cuda_time_us": 27263.346, + "pct_cuda_time": 98.26510213414547, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", - "cpu_time_us": 58.998, - "cuda_time_us": 43.384, - "pct_cuda_time": 0.1597150574992982, + "cpu_time_us": 53.801, + "cuda_time_us": 44.466, + "pct_cuda_time": 0.1602685169860263, "trace": "" }, "children": [ @@ -527,8 +527,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::indexSelectLargeIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, unsigned int, long)", "cpu_time_us": 0, - "cuda_time_us": 43.384, - "pct_cuda_time": 0.1597150574992982, + "cuda_time_us": 44.466, + "pct_cuda_time": 0.1602685169860263, "trace": "index_select(bfloat16[128256, 4096], 0, int64[512]) <- embedding(bfloat16[128256, 4096], int64[512], -1, False, False)" }, "children": [] @@ -538,18 +538,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1517.535, - "cuda_time_us": 912.268, - "pct_cuda_time": 3.358448646385068, + "cpu_time_us": 1635.054, + "cuda_time_us": 926.832, + "pct_cuda_time": 3.3405745993611458, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 79.279, - "cuda_time_us": 6.449, - "pct_cuda_time": 0.023741526964156694, + "cpu_time_us": 74.601, + "cuda_time_us": 6.929, + "pct_cuda_time": 0.02497415000666073, "trace": "" }, "children": [ @@ -557,8 +557,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, unsigned long, unsigned long)", "cpu_time_us": 0, - "cuda_time_us": 6.449, - "pct_cuda_time": 0.023741526964156694, + "cuda_time_us": 6.929, + "pct_cuda_time": 0.02497415000666073, "trace": "_C::rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -568,18 +568,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 1095.745, - "cuda_time_us": 311.696, - "pct_cuda_time": 1.1474862751775137, + "cpu_time_us": 1246.634, + "cuda_time_us": 332.68899999999996, + "pct_cuda_time": 1.199108816794047, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 197.775, - "cuda_time_us": 95.501, - "pct_cuda_time": 0.35158002273281574, + "cpu_time_us": 175.483, + "cuda_time_us": 95.892, + "pct_cuda_time": 0.34562291707875753, "trace": "" }, "children": [ @@ -587,8 +587,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 95.501, - "pct_cuda_time": 0.35158002273281574, + "cuda_time_us": 95.892, + "pct_cuda_time": 0.34562291707875753, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -598,9 +598,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 103.185, - "cuda_time_us": 34.848, - "pct_cuda_time": 0.12829039101363507, + "cpu_time_us": 95.944, + "cuda_time_us": 38.29299999999999, + "pct_cuda_time": 0.13801921290302485, "trace": "" }, "children": [ @@ -608,8 +608,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 6.809, - "pct_cuda_time": 0.025066840920909127, + "cuda_time_us": 6.689, + "pct_cuda_time": 0.024109119554705388, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -618,8 +618,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.286, - "pct_cuda_time": 0.015778598940669186, + "cuda_time_us": 5.928, + "pct_cuda_time": 0.02136625216329699, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -628,8 +628,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 10.415, - "pct_cuda_time": 0.03834206905437928, + "cuda_time_us": 11.176, + "pct_cuda_time": 0.0402815847127205, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -638,8 +638,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 9.013, - "pct_cuda_time": 0.0331807074783601, + "cuda_time_us": 10.334, + "pct_cuda_time": 0.03724676954377717, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -648,8 +648,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.325, - "pct_cuda_time": 0.015922174619317372, + "cuda_time_us": 4.166, + "pct_cuda_time": 0.015015486928524842, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -659,9 +659,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 620.336, - "cuda_time_us": 127.708, - "pct_cuda_time": 0.4701477633026087, + "cpu_time_us": 800.667, + "cuda_time_us": 128.19, + "pct_cuda_time": 0.46203439015064784, "trace": "" }, "children": [ @@ -669,8 +669,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.413, - "pct_cuda_time": 0.03465327854141836, + "cuda_time_us": 10.575, + "pct_cuda_time": 0.038115404289282326, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -679,8 +679,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 118.295, - "pct_cuda_time": 0.43549448476119035, + "cuda_time_us": 117.615, + "pct_cuda_time": 0.42391898586136556, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -690,9 +690,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 117.396, - "cuda_time_us": 53.639, - "pct_cuda_time": 0.19746809812845417, + "cpu_time_us": 116.084, + "cuda_time_us": 70.314, + "pct_cuda_time": 0.25343229666161676, "trace": "" }, "children": [ @@ -700,8 +700,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 53.639, - "pct_cuda_time": 0.19746809812845417, + "cuda_time_us": 70.314, + "pct_cuda_time": 0.25343229666161676, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -713,9 +713,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 51.507, - "cuda_time_us": 13.339, - "pct_cuda_time": 0.049106563525335115, + "cpu_time_us": 45.689, + "cuda_time_us": 12.898, + "pct_cuda_time": 0.04648817820550009, "trace": "" }, "children": [ @@ -723,8 +723,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.339, - "pct_cuda_time": 0.049106563525335115, + "cuda_time_us": 12.898, + "pct_cuda_time": 0.04648817820550009, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -734,18 +734,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 263.303, - "cuda_time_us": 580.784, - "pct_cuda_time": 2.138114280718062, + "cpu_time_us": 240.549, + "cuda_time_us": 574.316, + "pct_cuda_time": 2.0700034543549384, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 99.899, - "cuda_time_us": 341.699, - "pct_cuda_time": 1.2579401491898556, + "cpu_time_us": 96.114, + "cuda_time_us": 338.786, + "pct_cuda_time": 1.2210841945672628, "trace": "" }, "children": [ @@ -753,8 +753,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 341.699, - "pct_cuda_time": 1.2579401491898556, + "cuda_time_us": 338.786, + "pct_cuda_time": 1.2210841945672628, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -764,9 +764,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 36.544, - "cuda_time_us": 28.341, - "pct_cuda_time": 0.10433534124533493, + "cpu_time_us": 32.178, + "cuda_time_us": 28.481, + "pct_cuda_time": 0.10265388459225058, "trace": "" }, "children": [ @@ -774,8 +774,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 28.341, - "pct_cuda_time": 0.10433534124533493, + "cuda_time_us": 28.481, + "pct_cuda_time": 0.10265388459225058, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -785,9 +785,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 100.219, - "cuda_time_us": 210.744, - "pct_cuda_time": 0.7758387902828716, + "cpu_time_us": 88.653, + "cuda_time_us": 207.049, + "pct_cuda_time": 0.7462653751954248, "trace": "" }, "children": [ @@ -795,8 +795,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 210.744, - "pct_cuda_time": 0.7758387902828716, + "cuda_time_us": 207.049, + "pct_cuda_time": 0.7462653751954248, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -810,18 +810,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1113.792, - "cuda_time_us": 905.227, - "pct_cuda_time": 3.3325277142475844, + "cpu_time_us": 1083.249, + "cuda_time_us": 931.287, + "pct_cuda_time": 3.356631727125567, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.938, - "cuda_time_us": 14.621, - "pct_cuda_time": 0.05382615378243682, + "cpu_time_us": 35.263, + "cuda_time_us": 13.339, + "pct_cuda_time": 0.04807767166096804, "trace": "" }, "children": [ @@ -829,8 +829,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.621, - "pct_cuda_time": 0.05382615378243682, + "cuda_time_us": 13.339, + "pct_cuda_time": 0.04807767166096804, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -840,18 +840,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 781.206, - "cuda_time_us": 302.953, - "pct_cuda_time": 1.1152995531667178, + "cpu_time_us": 761.909, + "cuda_time_us": 327.269, + "pct_cuda_time": 1.1795735457540555, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 90.715, - "cuda_time_us": 91.796, - "pct_cuda_time": 0.3379403332612387, + "cpu_time_us": 86.109, + "cuda_time_us": 93.879, + "pct_cuda_time": 0.33836747416298213, "trace": "" }, "children": [ @@ -859,8 +859,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 91.796, - "pct_cuda_time": 0.3379403332612387, + "cuda_time_us": 93.879, + "pct_cuda_time": 0.33836747416298213, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -870,9 +870,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 92.568, - "cuda_time_us": 33.405, - "pct_cuda_time": 0.12297809090365242, + "cpu_time_us": 78.738, + "cuda_time_us": 33.205, + "pct_cuda_time": 0.1196805673215716, "trace": "" }, "children": [ @@ -880,8 +880,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 6.448, - "pct_cuda_time": 0.023737845536499054, + "cuda_time_us": 6.609, + "pct_cuda_time": 0.023820776070720272, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -891,7 +891,7 @@ "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 4.606, - "pct_cuda_time": 0.016956655791115792, + "pct_cuda_time": 0.01660137609044297, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -900,8 +900,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 9.974, - "pct_cuda_time": 0.03671855945735756, + "cuda_time_us": 9.733, + "pct_cuda_time": 0.035080589120339, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -910,8 +910,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.49, - "pct_cuda_time": 0.027573893155765805, + "cuda_time_us": 7.531, + "pct_cuda_time": 0.02714393472364872, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -920,8 +920,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.887, - "pct_cuda_time": 0.017991136962914215, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.017033891316420642, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -931,9 +931,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 454.239, - "cuda_time_us": 126.386, - "pct_cuda_time": 0.4652809159392012, + "cpu_time_us": 451.946, + "cuda_time_us": 127.668, + "pct_cuda_time": 0.460152948917645, "trace": "" }, "children": [ @@ -941,8 +941,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 8.972, - "pct_cuda_time": 0.03302976894439663, + "cuda_time_us": 9.132, + "pct_cuda_time": 0.03291440869690082, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -951,8 +951,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 117.414, - "pct_cuda_time": 0.4322511469948045, + "cuda_time_us": 118.536, + "pct_cuda_time": 0.4272385402207442, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -962,9 +962,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 105.919, - "cuda_time_us": 51.366, - "pct_cuda_time": 0.18910021306262567, + "cpu_time_us": 103.795, + "cuda_time_us": 72.517, + "pct_cuda_time": 0.2613725553518569, "trace": "" }, "children": [ @@ -972,8 +972,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 51.366, - "pct_cuda_time": 0.18910021306262567, + "cuda_time_us": 72.517, + "pct_cuda_time": 0.2613725553518569, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -985,9 +985,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 43.915, - "cuda_time_us": 12.698, - "pct_cuda_time": 0.04674676839678427, + "cpu_time_us": 41.572, + "cuda_time_us": 13.7, + "pct_cuda_time": 0.049378821632450866, "trace": "" }, "children": [ @@ -995,8 +995,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 12.698, - "pct_cuda_time": 0.04674676839678427, + "cuda_time_us": 13.7, + "pct_cuda_time": 0.049378821632450866, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1006,18 +1006,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 223.594, - "cuda_time_us": 574.955, - "pct_cuda_time": 2.1166552389016458, + "cpu_time_us": 219.789, + "cuda_time_us": 576.979, + "pct_cuda_time": 2.079601688078093, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 89.514, - "cuda_time_us": 341.869, - "pct_cuda_time": 1.2585659918916554, + "cpu_time_us": 87.331, + "cuda_time_us": 340.478, + "pct_cuda_time": 1.2271826592535477, "trace": "" }, "children": [ @@ -1025,8 +1025,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 341.869, - "pct_cuda_time": 1.2585659918916554, + "cuda_time_us": 340.478, + "pct_cuda_time": 1.2271826592535477, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -1036,9 +1036,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.64, - "cuda_time_us": 28.642, - "pct_cuda_time": 0.10544345097028625, + "cpu_time_us": 27.181, + "cuda_time_us": 27.229, + "pct_cuda_time": 0.09814130906788356, "trace": "" }, "children": [ @@ -1046,8 +1046,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 28.642, - "pct_cuda_time": 0.10544345097028625, + "cuda_time_us": 27.229, + "pct_cuda_time": 0.09814130906788356, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -1057,9 +1057,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 88.763, - "cuda_time_us": 204.444, - "pct_cuda_time": 0.752645796039704, + "cpu_time_us": 87.621, + "cuda_time_us": 209.272, + "pct_cuda_time": 0.7542777197566611, "trace": "" }, "children": [ @@ -1067,8 +1067,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 204.444, - "pct_cuda_time": 0.752645796039704, + "cuda_time_us": 209.272, + "pct_cuda_time": 0.7542777197566611, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -1082,18 +1082,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1077.739, - "cuda_time_us": 890.1949999999999, - "pct_cuda_time": 3.2771884936978553, + "cpu_time_us": 1080.034, + "cuda_time_us": 920.463, + "pct_cuda_time": 3.3176188537423807, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.263, - "cuda_time_us": 14.941, - "pct_cuda_time": 0.055004210632883424, + "cpu_time_us": 33.751, + "cuda_time_us": 14.621, + "pct_cuda_time": 0.05269837599182949, "trace": "" }, "children": [ @@ -1101,8 +1101,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.941, - "pct_cuda_time": 0.055004210632883424, + "cuda_time_us": 14.621, + "pct_cuda_time": 0.05269837599182949, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1112,18 +1112,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 754.726, - "cuda_time_us": 304.707, - "pct_cuda_time": 1.1217567772782284, + "cpu_time_us": 771.714, + "cuda_time_us": 328.933, + "pct_cuda_time": 1.185571090220946, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 83.175, - "cuda_time_us": 91.576, - "pct_cuda_time": 0.3371304191765566, + "cpu_time_us": 84.376, + "cuda_time_us": 95.632, + "pct_cuda_time": 0.34468580075580596, "trace": "" }, "children": [ @@ -1131,8 +1131,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 91.576, - "pct_cuda_time": 0.3371304191765566, + "cuda_time_us": 95.632, + "pct_cuda_time": 0.34468580075580596, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -1142,9 +1142,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 74.511, - "cuda_time_us": 34.126, - "pct_cuda_time": 0.12563240024481492, + "cpu_time_us": 77.566, + "cuda_time_us": 35.007, + "pct_cuda_time": 0.1261755042983363, "trace": "" }, "children": [ @@ -1152,8 +1152,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 6.729, - "pct_cuda_time": 0.024772326708297477, + "cuda_time_us": 6.889, + "pct_cuda_time": 0.02482997826466818, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -1162,8 +1162,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.016956655791115792, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.016889719574428082, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -1172,8 +1172,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 10.495, - "pct_cuda_time": 0.03863658326699093, + "cuda_time_us": 10.255, + "pct_cuda_time": 0.03696203035334187, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -1182,8 +1182,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.49, - "pct_cuda_time": 0.027573893155765805, + "cuda_time_us": 8.371, + "pct_cuda_time": 0.030171541305492423, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -1193,7 +1193,7 @@ "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 4.806, - "pct_cuda_time": 0.017692941322644918, + "pct_cuda_time": 0.017322234800405758, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -1203,9 +1203,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 457.583, - "cuda_time_us": 125.986, - "pct_cuda_time": 0.4638083448761429, + "cpu_time_us": 447.048, + "cuda_time_us": 126.708, + "pct_cuda_time": 0.4566928271098236, "trace": "" }, "children": [ @@ -1213,8 +1213,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 8.812, - "pct_cuda_time": 0.032440740519173325, + "cuda_time_us": 8.973, + "pct_cuda_time": 0.03234132602248041, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -1223,8 +1223,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 117.174, - "pct_cuda_time": 0.43136760435696964, + "cuda_time_us": 117.735, + "pct_cuda_time": 0.4243515010873432, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -1234,9 +1234,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 103.224, - "cuda_time_us": 53.019, - "pct_cuda_time": 0.19518561298071388, + "cpu_time_us": 122.964, + "cuda_time_us": 71.586, + "pct_cuda_time": 0.2580169580569801, "trace": "" }, "children": [ @@ -1244,8 +1244,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 53.019, - "pct_cuda_time": 0.19518561298071388, + "cuda_time_us": 71.586, + "pct_cuda_time": 0.2580169580569801, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -1257,9 +1257,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.509, - "cuda_time_us": 14.42, - "pct_cuda_time": 0.05308618682325005, + "cpu_time_us": 39.869, + "cuda_time_us": 12.979, + "pct_cuda_time": 0.04678012598303501, "trace": "" }, "children": [ @@ -1267,8 +1267,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.42, - "pct_cuda_time": 0.05308618682325005, + "cuda_time_us": 12.979, + "pct_cuda_time": 0.04678012598303501, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1278,18 +1278,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 224.736, - "cuda_time_us": 556.127, - "pct_cuda_time": 2.0473413189634937, + "cpu_time_us": 211.136, + "cuda_time_us": 563.9300000000001, + "pct_cuda_time": 2.0325692615465707, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 88.472, - "cuda_time_us": 331.063, - "pct_cuda_time": 1.2187844846231366, + "cpu_time_us": 85.157, + "cuda_time_us": 329.832, + "pct_cuda_time": 1.1888113501222286, "trace": "" }, "children": [ @@ -1297,8 +1297,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 331.063, - "pct_cuda_time": 1.2187844846231366, + "cuda_time_us": 329.832, + "pct_cuda_time": 1.1888113501222286, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -1308,9 +1308,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 27.161, - "cuda_time_us": 27.13, - "pct_cuda_time": 0.09987713235192605, + "cpu_time_us": 26.931, + "cuda_time_us": 27.87, + "pct_cuda_time": 0.10045166123331427, "trace": "" }, "children": [ @@ -1318,8 +1318,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 27.13, - "pct_cuda_time": 0.09987713235192605, + "cuda_time_us": 27.87, + "pct_cuda_time": 0.10045166123331427, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -1329,9 +1329,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 90.596, - "cuda_time_us": 197.934, - "pct_cuda_time": 0.728679701988431, + "cpu_time_us": 82.073, + "cuda_time_us": 206.228, + "pct_cuda_time": 0.7433062501910276, "trace": "" }, "children": [ @@ -1339,8 +1339,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 197.934, - "pct_cuda_time": 0.728679701988431, + "cuda_time_us": 206.228, + "pct_cuda_time": 0.7433062501910276, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -1354,18 +1354,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1060.994, - "cuda_time_us": 878.699, - "pct_cuda_time": 3.234866801345561, + "cpu_time_us": 1045.013, + "cuda_time_us": 909.766, + "pct_cuda_time": 3.2790637256400212, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.062, - "cuda_time_us": 13.459, - "pct_cuda_time": 0.04954833484425259, + "cpu_time_us": 33.55, + "cuda_time_us": 13.259, + "pct_cuda_time": 0.04778932817698292, "trace": "" }, "children": [ @@ -1373,8 +1373,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.459, - "pct_cuda_time": 0.04954833484425259, + "cuda_time_us": 13.259, + "pct_cuda_time": 0.04778932817698292, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1384,18 +1384,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 744.341, - "cuda_time_us": 304.537, - "pct_cuda_time": 1.1211309345764284, + "cpu_time_us": 730.061, + "cuda_time_us": 317.536, + "pct_cuda_time": 1.1444929566337165, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 86.258, - "cuda_time_us": 88.491, - "pct_cuda_time": 0.3257732148527199, + "cpu_time_us": 82.794, + "cuda_time_us": 93.578, + "pct_cuda_time": 0.3372825818044881, "trace": "" }, "children": [ @@ -1403,8 +1403,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 88.491, - "pct_cuda_time": 0.3257732148527199, + "cuda_time_us": 93.578, + "pct_cuda_time": 0.3372825818044881, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -1414,9 +1414,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 72.559, - "cuda_time_us": 35.769, - "pct_cuda_time": 0.13168098588632668, + "cpu_time_us": 74.932, + "cuda_time_us": 34.527, + "pct_cuda_time": 0.12444544339442562, "trace": "" }, "children": [ @@ -1424,8 +1424,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.05, - "pct_cuda_time": 0.02595406498640172, + "cuda_time_us": 6.689, + "pct_cuda_time": 0.024109119554705388, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -1434,8 +1434,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.01710391289742162, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.01645720434845041, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -1445,7 +1445,7 @@ "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.695, - "pct_cuda_time": 0.03937286879852006, + "pct_cuda_time": 0.03854791951526, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -1454,8 +1454,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.331, - "pct_cuda_time": 0.030669973815845776, + "cuda_time_us": 7.731, + "pct_cuda_time": 0.027864793433611507, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -1464,8 +1464,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.047, - "pct_cuda_time": 0.018580165388137514, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.017466406542398314, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -1475,9 +1475,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 445.094, - "cuda_time_us": 130.19299999999998, - "pct_cuda_time": 0.479296111031858, + "cpu_time_us": 434.54, + "cuda_time_us": 123.063, + "pct_cuda_time": 0.44355517712075193, "trace": "" }, "children": [ @@ -1485,8 +1485,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.173, - "pct_cuda_time": 0.033769735903583405, + "cuda_time_us": 9.133, + "pct_cuda_time": 0.032918012990450635, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -1495,8 +1495,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 121.02, - "pct_cuda_time": 0.4455263751282747, + "cuda_time_us": 113.93, + "pct_cuda_time": 0.41063716413030127, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -1506,9 +1506,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 103.915, - "cuda_time_us": 50.084, - "pct_cuda_time": 0.18438062280552397, + "cpu_time_us": 102.002, + "cuda_time_us": 66.368, + "pct_cuda_time": 0.239209754314051, "trace": "" }, "children": [ @@ -1516,8 +1516,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 50.084, - "pct_cuda_time": 0.18438062280552397, + "cuda_time_us": 66.368, + "pct_cuda_time": 0.239209754314051, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -1529,9 +1529,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.698, - "cuda_time_us": 14.461, - "pct_cuda_time": 0.05323712535721352, + "cpu_time_us": 39.088, + "cuda_time_us": 13.739, + "pct_cuda_time": 0.04951938908089361, "trace": "" }, "children": [ @@ -1539,8 +1539,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.461, - "pct_cuda_time": 0.05323712535721352, + "cuda_time_us": 13.739, + "pct_cuda_time": 0.04951938908089361, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1550,18 +1550,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 222.032, - "cuda_time_us": 546.242, - "pct_cuda_time": 2.0109504065676664, + "cpu_time_us": 221.321, + "cuda_time_us": 565.232, + "pct_cuda_time": 2.0372620517484283, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 89.884, - "cuda_time_us": 319.676, - "pct_cuda_time": 1.1768640678855258, + "cpu_time_us": 88.352, + "cuda_time_us": 331.004, + "pct_cuda_time": 1.1930355821626109, "trace": "" }, "children": [ @@ -1569,8 +1569,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 319.676, - "pct_cuda_time": 1.1768640678855258, + "cuda_time_us": 331.004, + "pct_cuda_time": 1.1930355821626109, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -1580,9 +1580,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 29.595, - "cuda_time_us": 27.15, - "pct_cuda_time": 0.09995076090507898, + "cpu_time_us": 27.451, + "cuda_time_us": 27.841, + "pct_cuda_time": 0.10034713672036968, "trace": "" }, "children": [ @@ -1590,8 +1590,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 27.15, - "pct_cuda_time": 0.09995076090507898, + "cuda_time_us": 27.841, + "pct_cuda_time": 0.10034713672036968, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -1601,9 +1601,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 84.306, - "cuda_time_us": 199.416, - "pct_cuda_time": 0.7341355777770618, + "cpu_time_us": 83.284, + "cuda_time_us": 206.387, + "pct_cuda_time": 0.7438793328654479, "trace": "" }, "children": [ @@ -1611,8 +1611,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 199.416, - "pct_cuda_time": 0.7341355777770618, + "cuda_time_us": 206.387, + "pct_cuda_time": 0.7438793328654479, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -1626,18 +1626,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1081.254, - "cuda_time_us": 865.418, - "pct_cuda_time": 3.18597376062437, + "cpu_time_us": 1088.107, + "cuda_time_us": 921.595, + "pct_cuda_time": 3.3216989140407702, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.459, - "cuda_time_us": 14.501, - "pct_cuda_time": 0.05338438246351934, + "cpu_time_us": 50.606, + "cuda_time_us": 13.099, + "pct_cuda_time": 0.04721264120901269, "trace": "" }, "children": [ @@ -1645,8 +1645,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.501, - "pct_cuda_time": 0.05338438246351934, + "cuda_time_us": 13.099, + "pct_cuda_time": 0.04721264120901269, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1656,18 +1656,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 756.73, - "cuda_time_us": 304.545, - "pct_cuda_time": 1.1211603859976897, + "cpu_time_us": 759.085, + "cuda_time_us": 323.255, + "pct_cuda_time": 1.1651059114451026, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 81.842, - "cuda_time_us": 90.314, - "pct_cuda_time": 0.3324844574726078, + "cpu_time_us": 88.532, + "cuda_time_us": 91.906, + "pct_cuda_time": 0.3312562029891992, "trace": "" }, "children": [ @@ -1675,8 +1675,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 90.314, - "pct_cuda_time": 0.3324844574726078, + "cuda_time_us": 91.906, + "pct_cuda_time": 0.3312562029891992, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -1686,9 +1686,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 85.187, - "cuda_time_us": 34.205999999999996, - "pct_cuda_time": 0.12592691445742657, + "cpu_time_us": 75.493, + "cuda_time_us": 35.368, + "pct_cuda_time": 0.12747665426981913, "trace": "" }, "children": [ @@ -1696,8 +1696,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 6.889, - "pct_cuda_time": 0.025361355133520776, + "cuda_time_us": 6.689, + "pct_cuda_time": 0.024109119554705388, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -1707,7 +1707,7 @@ "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 4.686, - "pct_cuda_time": 0.017251170003727445, + "pct_cuda_time": 0.016889719574428082, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -1716,8 +1716,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 9.814, - "pct_cuda_time": 0.03612953103213425, + "cuda_time_us": 11.296, + "pct_cuda_time": 0.04071409993869817, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -1726,8 +1726,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.81, - "pct_cuda_time": 0.028751950006212404, + "cuda_time_us": 7.971, + "pct_cuda_time": 0.02872982388556685, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -1736,8 +1736,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.007, - "pct_cuda_time": 0.01843290828183169, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.017033891316420642, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -1747,9 +1747,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 448.75, - "cuda_time_us": 130.75300000000001, - "pct_cuda_time": 0.48135771052013976, + "cpu_time_us": 454.008, + "cuda_time_us": 125.226, + "pct_cuda_time": 0.4513512640689994, "trace": "" }, "children": [ @@ -1757,8 +1757,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.493, - "pct_cuda_time": 0.03494779275403001, + "cuda_time_us": 8.933, + "pct_cuda_time": 0.032197154280487854, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -1767,8 +1767,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 121.26, - "pct_cuda_time": 0.44640991776610967, + "cuda_time_us": 116.293, + "pct_cuda_time": 0.4191541097885116, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -1778,9 +1778,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 104.937, - "cuda_time_us": 49.272, - "pct_cuda_time": 0.1813913035475157, + "cpu_time_us": 105.318, + "cuda_time_us": 70.755, + "pct_cuda_time": 0.2550217901170847, "trace": "" }, "children": [ @@ -1788,8 +1788,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 49.272, - "pct_cuda_time": 0.1813913035475157, + "cuda_time_us": 70.755, + "pct_cuda_time": 0.2550217901170847, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -1801,9 +1801,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 47.501, - "cuda_time_us": 13.66, - "pct_cuda_time": 0.05028830180343937, + "cpu_time_us": 39.228, + "cuda_time_us": 13.539, + "pct_cuda_time": 0.04879853037093082, "trace": "" }, "children": [ @@ -1811,8 +1811,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.66, - "pct_cuda_time": 0.05028830180343937, + "cuda_time_us": 13.539, + "pct_cuda_time": 0.04879853037093082, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1822,18 +1822,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 223.224, - "cuda_time_us": 532.712, - "pct_cuda_time": 1.9611406903597213, + "cpu_time_us": 215.132, + "cuda_time_us": 571.702, + "pct_cuda_time": 2.0605818310157247, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 86.178, - "cuda_time_us": 312.946, - "pct_cuda_time": 1.1520880597495708, + "cpu_time_us": 86.749, + "cuda_time_us": 336.252, + "pct_cuda_time": 1.2119509147120342, "trace": "" }, "children": [ @@ -1841,8 +1841,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 312.946, - "pct_cuda_time": 1.1520880597495708, + "cuda_time_us": 336.252, + "pct_cuda_time": 1.2119509147120342, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -1852,9 +1852,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 30.586, - "cuda_time_us": 28.211, - "pct_cuda_time": 0.10385675564984098, + "cpu_time_us": 26.89, + "cuda_time_us": 28.602, + "pct_cuda_time": 0.10309000411177809, "trace": "" }, "children": [ @@ -1862,8 +1862,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 28.211, - "pct_cuda_time": 0.10385675564984098, + "cuda_time_us": 28.602, + "pct_cuda_time": 0.10309000411177809, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -1873,9 +1873,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 88.662, - "cuda_time_us": 191.555, - "pct_cuda_time": 0.7051958749603096, + "cpu_time_us": 83.775, + "cuda_time_us": 206.848, + "pct_cuda_time": 0.7455409121919122, "trace": "" }, "children": [ @@ -1883,8 +1883,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 191.555, - "pct_cuda_time": 0.7051958749603096, + "cuda_time_us": 206.848, + "pct_cuda_time": 0.7455409121919122, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -1898,18 +1898,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1069.877, - "cuda_time_us": 883.9449999999999, - "pct_cuda_time": 3.2541795708375703, + "cpu_time_us": 1044.422, + "cuda_time_us": 906.5520000000001, + "pct_cuda_time": 3.2674795261709195, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.383, - "cuda_time_us": 15.021, - "pct_cuda_time": 0.05529872484549508, + "cpu_time_us": 31.747, + "cuda_time_us": 13.819, + "pct_cuda_time": 0.04980773256487872, "trace": "" }, "children": [ @@ -1917,8 +1917,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.021, - "pct_cuda_time": 0.05529872484549508, + "cuda_time_us": 13.819, + "pct_cuda_time": 0.04980773256487872, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1928,18 +1928,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 749.038, - "cuda_time_us": 308.361, - "pct_cuda_time": 1.1352087139392655, + "cpu_time_us": 735.93, + "cuda_time_us": 320.972, + "pct_cuda_time": 1.1568773092708773, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.096, - "cuda_time_us": 86.619, - "pct_cuda_time": 0.3188815822776072, + "cpu_time_us": 82.664, + "cuda_time_us": 91.206, + "pct_cuda_time": 0.32873319750432944, "trace": "" }, "children": [ @@ -1947,8 +1947,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 86.619, - "pct_cuda_time": 0.3188815822776072, + "cuda_time_us": 91.206, + "pct_cuda_time": 0.32873319750432944, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -1958,9 +1958,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 86.52, - "cuda_time_us": 36.409, - "pct_cuda_time": 0.1340370995872199, + "cpu_time_us": 75.683, + "cuda_time_us": 35.888, + "pct_cuda_time": 0.12935088691572236, "trace": "" }, "children": [ @@ -1968,8 +1968,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.009, - "pct_cuda_time": 0.025803126452438253, + "cuda_time_us": 6.809, + "pct_cuda_time": 0.024541634780683063, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -1978,8 +1978,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.806, - "pct_cuda_time": 0.017692941322644918, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.017178063058413198, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -1988,8 +1988,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.136, - "pct_cuda_time": 0.04099637839554179, + "cuda_time_us": 11.056, + "pct_cuda_time": 0.03984906948674282, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -1998,8 +1998,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.211, - "pct_cuda_time": 0.030228202496928307, + "cuda_time_us": 8.491, + "pct_cuda_time": 0.03060405653147009, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -2008,8 +2008,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.247, - "pct_cuda_time": 0.019316450919666644, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.017178063058413198, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -2019,9 +2019,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 443.272, - "cuda_time_us": 137.082, - "pct_cuda_time": 0.5046574661653788, + "cpu_time_us": 440.849, + "cuda_time_us": 127.149, + "pct_cuda_time": 0.4582823205652916, "trace": "" }, "children": [ @@ -2029,8 +2029,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.693, - "pct_cuda_time": 0.03568407828555913, + "cuda_time_us": 8.893, + "pct_cuda_time": 0.0320529825384953, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -2039,8 +2039,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 127.389, - "pct_cuda_time": 0.46897338787981974, + "cuda_time_us": 118.256, + "pct_cuda_time": 0.4262293380267963, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -2050,9 +2050,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.037, - "cuda_time_us": 48.251, - "pct_cuda_time": 0.17763256590905951, + "cpu_time_us": 101.341, + "cuda_time_us": 66.729, + "pct_cuda_time": 0.24051090428553384, "trace": "" }, "children": [ @@ -2060,8 +2060,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 48.251, - "pct_cuda_time": 0.17763256590905951, + "cuda_time_us": 66.729, + "pct_cuda_time": 0.24051090428553384, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -2073,9 +2073,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 40.69, - "cuda_time_us": 14.02, - "pct_cuda_time": 0.051613615760191786, + "cpu_time_us": 39.689, + "cuda_time_us": 14.1, + "pct_cuda_time": 0.050820539052376434, "trace": "" }, "children": [ @@ -2083,8 +2083,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.02, - "pct_cuda_time": 0.051613615760191786, + "cuda_time_us": 14.1, + "pct_cuda_time": 0.050820539052376434, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2094,18 +2094,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 220.109, - "cuda_time_us": 546.543, - "pct_cuda_time": 2.012058516292618, + "cpu_time_us": 213.169, + "cuda_time_us": 557.6610000000001, + "pct_cuda_time": 2.009973945282787, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 88.131, - "cuda_time_us": 318.925, - "pct_cuda_time": 1.174099315714634, + "cpu_time_us": 86.91, + "cuda_time_us": 325.986, + "pct_cuda_time": 1.1749492371296444, "trace": "" }, "children": [ @@ -2113,8 +2113,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 318.925, - "pct_cuda_time": 1.174099315714634, + "cuda_time_us": 325.986, + "pct_cuda_time": 1.1749492371296444, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -2124,9 +2124,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.479, - "cuda_time_us": 28.482, - "pct_cuda_time": 0.10485442254506296, + "cpu_time_us": 26.049, + "cuda_time_us": 27.841, + "pct_cuda_time": 0.10034713672036968, "trace": "" }, "children": [ @@ -2134,8 +2134,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 28.482, - "pct_cuda_time": 0.10485442254506296, + "cuda_time_us": 27.841, + "pct_cuda_time": 0.10034713672036968, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -2145,9 +2145,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 88.202, - "cuda_time_us": 199.136, - "pct_cuda_time": 0.7331047780329211, + "cpu_time_us": 81.692, + "cuda_time_us": 203.834, + "pct_cuda_time": 0.734677571432773, "trace": "" }, "children": [ @@ -2155,8 +2155,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 199.136, - "pct_cuda_time": 0.7331047780329211, + "cuda_time_us": 203.834, + "pct_cuda_time": 0.734677571432773, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -2170,18 +2170,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1087.053, - "cuda_time_us": 878.978, - "pct_cuda_time": 3.2358939196620446, + "cpu_time_us": 1071.622, + "cuda_time_us": 902.615, + "pct_cuda_time": 3.253289422465302, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 34.973, - "cuda_time_us": 13.739, - "pct_cuda_time": 0.050579134588393374, + "cpu_time_us": 33.19, + "cuda_time_us": 14.5, + "pct_cuda_time": 0.05226225647230201, "trace": "" }, "children": [ @@ -2189,8 +2189,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.739, - "pct_cuda_time": 0.050579134588393374, + "cuda_time_us": 14.5, + "pct_cuda_time": 0.05226225647230201, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2200,18 +2200,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 766.355, - "cuda_time_us": 306.249, - "pct_cuda_time": 1.127433538726318, + "cpu_time_us": 763.391, + "cuda_time_us": 321.262, + "pct_cuda_time": 1.1579225544003233, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 104.957, - "cuda_time_us": 88.852, - "pct_cuda_time": 0.32710221023712993, + "cpu_time_us": 79.979, + "cuda_time_us": 89.693, + "pct_cuda_time": 0.32327990136346096, "trace": "" }, "children": [ @@ -2219,8 +2219,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 88.852, - "pct_cuda_time": 0.32710221023712993, + "cuda_time_us": 89.693, + "pct_cuda_time": 0.32327990136346096, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -2230,9 +2230,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 80.57, - "cuda_time_us": 34.767, - "pct_cuda_time": 0.12799219537336579, + "cpu_time_us": 82.263, + "cuda_time_us": 34.407, + "pct_cuda_time": 0.12401292816844793, "trace": "" }, "children": [ @@ -2240,8 +2240,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.09, - "pct_cuda_time": 0.02610132209270755, + "cuda_time_us": 6.93, + "pct_cuda_time": 0.02497775430021055, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -2250,8 +2250,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.525, - "pct_cuda_time": 0.0166584601508465, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.017033891316420642, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -2260,8 +2260,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 10.375, - "pct_cuda_time": 0.038194811948073455, + "cuda_time_us": 10.294, + "pct_cuda_time": 0.03710259780178462, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -2270,8 +2270,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.73, - "pct_cuda_time": 0.028457435793600754, + "cuda_time_us": 7.771, + "pct_cuda_time": 0.028008965175604066, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -2280,8 +2280,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.047, - "pct_cuda_time": 0.018580165388137514, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.016889719574428082, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -2291,9 +2291,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 443.903, - "cuda_time_us": 132.596, - "pct_cuda_time": 0.4881425816931806, + "cpu_time_us": 462.612, + "cuda_time_us": 127.869, + "pct_cuda_time": 0.4608774119211576, "trace": "" }, "children": [ @@ -2301,8 +2301,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.293, - "pct_cuda_time": 0.03421150722250088, + "cuda_time_us": 9.813, + "pct_cuda_time": 0.03536893260432412, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -2311,8 +2311,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 123.303, - "pct_cuda_time": 0.45393107447067965, + "cuda_time_us": 118.056, + "pct_cuda_time": 0.4255084793168335, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -2322,9 +2322,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 100.39, - "cuda_time_us": 50.034, - "pct_cuda_time": 0.18419655142264169, + "cpu_time_us": 103.074, + "cuda_time_us": 69.293, + "pct_cuda_time": 0.24975231294725678, "trace": "" }, "children": [ @@ -2332,8 +2332,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 50.034, - "pct_cuda_time": 0.18419655142264169, + "cuda_time_us": 69.293, + "pct_cuda_time": 0.24975231294725678, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -2345,9 +2345,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 40.28, - "cuda_time_us": 13.138, - "pct_cuda_time": 0.04836659656614834, + "cpu_time_us": 38.898, + "cuda_time_us": 12.858, + "pct_cuda_time": 0.04634400646350754, "trace": "" }, "children": [ @@ -2355,8 +2355,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.138, - "pct_cuda_time": 0.04836659656614834, + "cuda_time_us": 12.858, + "pct_cuda_time": 0.04634400646350754, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2366,18 +2366,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 224.986, - "cuda_time_us": 545.852, - "pct_cuda_time": 2.0095146497811847, + "cpu_time_us": 215.071, + "cuda_time_us": 553.995, + "pct_cuda_time": 1.996760605129169, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 90.595, - "cuda_time_us": 317.233, - "pct_cuda_time": 1.1678703401178976, + "cpu_time_us": 89.714, + "cuda_time_us": 329.151, + "pct_cuda_time": 1.1863568262148054, "trace": "" }, "children": [ @@ -2385,8 +2385,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 317.233, - "pct_cuda_time": 1.1678703401178976, + "cuda_time_us": 329.151, + "pct_cuda_time": 1.1863568262148054, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -2396,9 +2396,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 29.054, - "cuda_time_us": 27.209, - "pct_cuda_time": 0.10016796513688006, + "cpu_time_us": 26.77, + "cuda_time_us": 26.879, + "pct_cuda_time": 0.09687980632544867, "trace": "" }, "children": [ @@ -2406,8 +2406,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 27.209, - "pct_cuda_time": 0.10016796513688006, + "cuda_time_us": 26.879, + "pct_cuda_time": 0.09687980632544867, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -2417,9 +2417,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 87.341, - "cuda_time_us": 201.41, - "pct_cuda_time": 0.7414763445264072, + "cpu_time_us": 81.031, + "cuda_time_us": 197.965, + "pct_cuda_time": 0.713523972588915, "trace": "" }, "children": [ @@ -2427,8 +2427,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 201.41, - "pct_cuda_time": 0.7414763445264072, + "cuda_time_us": 197.965, + "pct_cuda_time": 0.713523972588915, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -2442,18 +2442,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1044.349, - "cuda_time_us": 897.2959999999999, - "pct_cuda_time": 3.303330311494797, + "cpu_time_us": 1024.411, + "cuda_time_us": 901.604, + "pct_cuda_time": 3.24964548168644, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.74, - "cuda_time_us": 13.419, - "pct_cuda_time": 0.04940107773794677, + "cpu_time_us": 34.011, + "cuda_time_us": 14.381, + "pct_cuda_time": 0.051833345539874155, "trace": "" }, "children": [ @@ -2461,8 +2461,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.419, - "pct_cuda_time": 0.04940107773794677, + "cuda_time_us": 14.381, + "pct_cuda_time": 0.051833345539874155, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2472,18 +2472,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 730.972, - "cuda_time_us": 317.685, - "pct_cuda_time": 1.1695343454191534, + "cpu_time_us": 722.66, + "cuda_time_us": 327.26, + "pct_cuda_time": 1.1795411071121074, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.115, - "cuda_time_us": 88.071, - "pct_cuda_time": 0.32422701523650865, + "cpu_time_us": 82.133, + "cuda_time_us": 88.982, + "pct_cuda_time": 0.32071724864954326, "trace": "" }, "children": [ @@ -2491,8 +2491,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 88.071, - "pct_cuda_time": 0.32422701523650865, + "cuda_time_us": 88.982, + "pct_cuda_time": 0.32071724864954326, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -2502,9 +2502,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 76.915, - "cuda_time_us": 34.726, - "pct_cuda_time": 0.1278412568394023, + "cpu_time_us": 74.401, + "cuda_time_us": 34.887, + "pct_cuda_time": 0.12574298907235862, "trace": "" }, "children": [ @@ -2513,7 +2513,7 @@ "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 6.889, - "pct_cuda_time": 0.025361355133520776, + "pct_cuda_time": 0.02482997826466818, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -2522,8 +2522,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.926, - "pct_cuda_time": 0.018134712641562394, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.01645720434845041, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -2532,8 +2532,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 10.255, - "pct_cuda_time": 0.037753040629155986, + "cuda_time_us": 11.055, + "pct_cuda_time": 0.039845465193193014, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -2542,8 +2542,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.69, - "pct_cuda_time": 0.02831017868729493, + "cuda_time_us": 7.611, + "pct_cuda_time": 0.027432278207633835, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -2552,8 +2552,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.966, - "pct_cuda_time": 0.01828196974786822, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.017178063058413198, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -2563,9 +2563,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 431.685, - "cuda_time_us": 130.794, - "pct_cuda_time": 0.4815086490541032, + "cpu_time_us": 428.941, + "cuda_time_us": 128.39, + "pct_cuda_time": 0.4627552488606106, "trace": "" }, "children": [ @@ -2573,8 +2573,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.574, - "pct_cuda_time": 0.0352459883942993, + "cuda_time_us": 9.293, + "pct_cuda_time": 0.03349469995842086, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -2583,8 +2583,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 121.22, - "pct_cuda_time": 0.4462626606598038, + "cuda_time_us": 119.097, + "pct_cuda_time": 0.42926054890218984, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -2594,9 +2594,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.509, - "cuda_time_us": 64.094, - "pct_cuda_time": 0.2359574242891393, + "cpu_time_us": 98.497, + "cuda_time_us": 75.001, + "pct_cuda_time": 0.2703256205295947, "trace": "" }, "children": [ @@ -2604,8 +2604,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 64.094, - "pct_cuda_time": 0.2359574242891393, + "cuda_time_us": 75.001, + "pct_cuda_time": 0.2703256205295947, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -2617,9 +2617,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.907, - "cuda_time_us": 13.219, - "pct_cuda_time": 0.04866479220641764, + "cpu_time_us": 37.786, + "cuda_time_us": 15.142, + "pct_cuda_time": 0.05457621293128255, "trace": "" }, "children": [ @@ -2627,8 +2627,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.219, - "pct_cuda_time": 0.04866479220641764, + "cuda_time_us": 15.142, + "pct_cuda_time": 0.05457621293128255, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2638,18 +2638,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 219.427, - "cuda_time_us": 552.973, - "pct_cuda_time": 2.035730096131279, + "cpu_time_us": 208.792, + "cuda_time_us": 544.821, + "pct_cuda_time": 1.9636948161031762, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 84.366, - "cuda_time_us": 324.824, - "pct_cuda_time": 1.1958160574670857, + "cpu_time_us": 84.086, + "cuda_time_us": 322.29, + "pct_cuda_time": 1.1616277681695322, "trace": "" }, "children": [ @@ -2657,8 +2657,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 324.824, - "pct_cuda_time": 1.1958160574670857, + "cuda_time_us": 322.29, + "pct_cuda_time": 1.1616277681695322, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -2668,9 +2668,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 29.184, - "cuda_time_us": 26.869, - "pct_cuda_time": 0.09891627973328054, + "cpu_time_us": 25.669, + "cuda_time_us": 27.18, + "pct_cuda_time": 0.09796469868394266, "trace": "" }, "children": [ @@ -2678,8 +2678,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 26.869, - "pct_cuda_time": 0.09891627973328054, + "cuda_time_us": 27.18, + "pct_cuda_time": 0.09796469868394266, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -2689,9 +2689,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 88.533, - "cuda_time_us": 201.28, - "pct_cuda_time": 0.7409977589309134, + "cpu_time_us": 82.193, + "cuda_time_us": 195.351, + "pct_cuda_time": 0.7041023492497014, "trace": "" }, "children": [ @@ -2699,8 +2699,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 201.28, - "pct_cuda_time": 0.7409977589309134, + "cuda_time_us": 195.351, + "pct_cuda_time": 0.7041023492497014, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -2714,18 +2714,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1060.012, - "cuda_time_us": 876.414, - "pct_cuda_time": 3.226454739147841, + "cpu_time_us": 1024.331, + "cuda_time_us": 873.14, + "pct_cuda_time": 3.1470528700845364, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 36.534, - "cuda_time_us": 13.579, - "pct_cuda_time": 0.04999010616317007, + "cpu_time_us": 32.889, + "cuda_time_us": 14.381, + "pct_cuda_time": 0.051833345539874155, "trace": "" }, "children": [ @@ -2733,8 +2733,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.579, - "pct_cuda_time": 0.04999010616317007, + "cuda_time_us": 14.381, + "pct_cuda_time": 0.051833345539874155, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2744,18 +2744,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 739.744, - "cuda_time_us": 301.952, - "pct_cuda_time": 1.1116144440814146, + "cpu_time_us": 721.218, + "cuda_time_us": 323.183, + "pct_cuda_time": 1.1648464023095157, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.246, - "cuda_time_us": 89.402, - "pct_cuda_time": 0.329126995448835, + "cpu_time_us": 80.57, + "cuda_time_us": 88.951, + "pct_cuda_time": 0.320605515549499, "trace": "" }, "children": [ @@ -2763,8 +2763,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 89.402, - "pct_cuda_time": 0.329126995448835, + "cuda_time_us": 88.951, + "pct_cuda_time": 0.320605515549499, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -2774,9 +2774,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 83.545, - "cuda_time_us": 33.806, - "pct_cuda_time": 0.12445434339436831, + "cpu_time_us": 74.291, + "cuda_time_us": 35.568999999999996, + "pct_cuda_time": 0.12820111727333172, "trace": "" }, "children": [ @@ -2784,8 +2784,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 6.77, - "pct_cuda_time": 0.024923265242260944, + "cuda_time_us": 7.09, + "pct_cuda_time": 0.025554441268180773, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -2794,8 +2794,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.645, - "pct_cuda_time": 0.01710023146976397, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.017033891316420642, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -2804,8 +2804,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 9.934, - "pct_cuda_time": 0.03657130235105173, + "cuda_time_us": 10.615, + "pct_cuda_time": 0.03825957603127489, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -2814,8 +2814,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.611, - "pct_cuda_time": 0.028019345902340922, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.029883197821507307, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -2824,8 +2824,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.846, - "pct_cuda_time": 0.017840198428950745, + "cuda_time_us": 4.847, + "pct_cuda_time": 0.017470010835948128, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -2835,9 +2835,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 434.408, - "cuda_time_us": 128.79, - "pct_cuda_time": 0.4741310680281812, + "cpu_time_us": 434.73, + "cuda_time_us": 132.996, + "pct_cuda_time": 0.47935662495105374, "trace": "" }, "children": [ @@ -2845,8 +2845,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.293, - "pct_cuda_time": 0.03421150722250088, + "cuda_time_us": 10.214, + "pct_cuda_time": 0.0368142543177995, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -2855,8 +2855,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 119.497, - "pct_cuda_time": 0.43991956080568034, + "cuda_time_us": 122.782, + "pct_cuda_time": 0.44254237063325413, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -2866,9 +2866,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 101.321, - "cuda_time_us": 49.954, - "pct_cuda_time": 0.18390203721003, + "cpu_time_us": 97.035, + "cuda_time_us": 65.667, + "pct_cuda_time": 0.23668314453563147, "trace": "" }, "children": [ @@ -2876,8 +2876,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 49.954, - "pct_cuda_time": 0.18390203721003, + "cuda_time_us": 65.667, + "pct_cuda_time": 0.23668314453563147, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -2889,9 +2889,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.509, - "cuda_time_us": 14.06, - "pct_cuda_time": 0.05176087286649762, + "cpu_time_us": 38.518, + "cuda_time_us": 13.779, + "pct_cuda_time": 0.049663560822886164, "trace": "" }, "children": [ @@ -2899,8 +2899,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.06, - "pct_cuda_time": 0.05176087286649762, + "cuda_time_us": 13.779, + "pct_cuda_time": 0.049663560822886164, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2910,18 +2910,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 222.212, - "cuda_time_us": 546.823, - "pct_cuda_time": 2.013089316036759, + "cpu_time_us": 211.466, + "cuda_time_us": 521.797, + "pct_cuda_time": 1.8807095614122602, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 89.874, - "cuda_time_us": 318.384, - "pct_cuda_time": 1.1721076633518477, + "cpu_time_us": 84.917, + "cuda_time_us": 303.894, + "pct_cuda_time": 1.095323184027155, "trace": "" }, "children": [ @@ -2929,8 +2929,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 318.384, - "pct_cuda_time": 1.1721076633518477, + "cuda_time_us": 303.894, + "pct_cuda_time": 1.095323184027155, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -2940,9 +2940,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 27.992, - "cuda_time_us": 27.2, - "pct_cuda_time": 0.10013483228796126, + "cpu_time_us": 26.75, + "cuda_time_us": 26.008, + "pct_cuda_time": 0.09374046664356073, "trace": "" }, "children": [ @@ -2950,8 +2950,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 27.2, - "pct_cuda_time": 0.10013483228796126, + "cuda_time_us": 26.008, + "pct_cuda_time": 0.09374046664356073, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -2961,9 +2961,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 86.8, - "cuda_time_us": 201.239, - "pct_cuda_time": 0.7408468203969499, + "cpu_time_us": 81.822, + "cuda_time_us": 191.895, + "pct_cuda_time": 0.6916459107415445, "trace": "" }, "children": [ @@ -2971,8 +2971,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 201.239, - "pct_cuda_time": 0.7408468203969499, + "cuda_time_us": 191.895, + "pct_cuda_time": 0.6916459107415445, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -2986,18 +2986,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1056.407, - "cuda_time_us": 868.4929999999999, - "pct_cuda_time": 3.1972941506716297, + "cpu_time_us": 1055.588, + "cuda_time_us": 869.996, + "pct_cuda_time": 3.135720971163921, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.163, - "cuda_time_us": 13.7, - "pct_cuda_time": 0.05043555890974519, + "cpu_time_us": 34.652, + "cuda_time_us": 13.46, + "pct_cuda_time": 0.04851379118049552, "trace": "" }, "children": [ @@ -3005,8 +3005,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.7, - "pct_cuda_time": 0.05043555890974519, + "cuda_time_us": 13.46, + "pct_cuda_time": 0.04851379118049552, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3016,18 +3016,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 733.305, - "cuda_time_us": 305.707, - "pct_cuda_time": 1.125438204935874, + "cpu_time_us": 741.939, + "cuda_time_us": 329.132, + "pct_cuda_time": 1.186288344637359, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 83.164, - "cuda_time_us": 90.714, - "pct_cuda_time": 0.33395702853566606, + "cpu_time_us": 82.914, + "cuda_time_us": 85.897, + "pct_cuda_time": 0.3095980030483673, "trace": "" }, "children": [ @@ -3035,8 +3035,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 90.714, - "pct_cuda_time": 0.33395702853566606, + "cuda_time_us": 85.897, + "pct_cuda_time": 0.3095980030483673, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -3046,9 +3046,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 80.14, - "cuda_time_us": 34.326, - "pct_cuda_time": 0.12636868577634405, + "cpu_time_us": 76.955, + "cuda_time_us": 36.129, + "pct_cuda_time": 0.13021952166122752, "trace": "" }, "children": [ @@ -3056,8 +3056,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.09, - "pct_cuda_time": 0.02610132209270755, + "cuda_time_us": 7.17, + "pct_cuda_time": 0.02584278475216589, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -3066,8 +3066,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.806, - "pct_cuda_time": 0.017692941322644918, + "cuda_time_us": 4.926, + "pct_cuda_time": 0.017754750026383426, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -3076,8 +3076,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 9.653, - "pct_cuda_time": 0.03553682117925331, + "cuda_time_us": 10.735, + "pct_cuda_time": 0.03869209125725255, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -3086,8 +3086,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.851, - "pct_cuda_time": 0.028902888540175878, + "cuda_time_us": 8.372, + "pct_cuda_time": 0.030175145599042233, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -3097,7 +3097,7 @@ "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 4.926, - "pct_cuda_time": 0.018134712641562394, + "pct_cuda_time": 0.017754750026383426, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -3107,9 +3107,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 436.752, - "cuda_time_us": 131.715, - "pct_cuda_time": 0.48489924392679473, + "cpu_time_us": 448.781, + "cuda_time_us": 134.358, + "pct_cuda_time": 0.4842656727659002, "trace": "" }, "children": [ @@ -3117,8 +3117,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 8.933, - "pct_cuda_time": 0.03288619326574845, + "cuda_time_us": 9.733, + "pct_cuda_time": 0.035080589120339, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -3127,8 +3127,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 122.782, - "pct_cuda_time": 0.45201305066104625, + "cuda_time_us": 124.625, + "pct_cuda_time": 0.44918508364556126, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -3138,9 +3138,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 97.746, - "cuda_time_us": 48.952, - "pct_cuda_time": 0.1802132466970691, + "cpu_time_us": 97.856, + "cuda_time_us": 72.748, + "pct_cuda_time": 0.26220514716186394, "trace": "" }, "children": [ @@ -3148,8 +3148,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 48.952, - "pct_cuda_time": 0.1802132466970691, + "cuda_time_us": 72.748, + "pct_cuda_time": 0.26220514716186394, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -3161,9 +3161,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.9, - "cuda_time_us": 13.74, - "pct_cuda_time": 0.050582816016051024, + "cpu_time_us": 39.84, + "cuda_time_us": 14.14, + "pct_cuda_time": 0.05096471079436899, "trace": "" }, "children": [ @@ -3171,8 +3171,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.74, - "pct_cuda_time": 0.050582816016051024, + "cuda_time_us": 14.14, + "pct_cuda_time": 0.05096471079436899, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3182,18 +3182,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 226.098, - "cuda_time_us": 535.346, - "pct_cuda_time": 1.97083757080996, + "cpu_time_us": 210.395, + "cuda_time_us": 513.264, + "pct_cuda_time": 1.849954124551698, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 94.29, - "cuda_time_us": 314.769, - "pct_cuda_time": 1.1587993023694587, + "cpu_time_us": 82.523, + "cuda_time_us": 298.856, + "pct_cuda_time": 1.0771647531231923, "trace": "" }, "children": [ @@ -3201,8 +3201,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 314.769, - "pct_cuda_time": 1.1587993023694587, + "cuda_time_us": 298.856, + "pct_cuda_time": 1.0771647531231923, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -3212,9 +3212,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 27.631, - "cuda_time_us": 27.139, - "pct_cuda_time": 0.09991026520084487, + "cpu_time_us": 26.3, + "cuda_time_us": 24.836, + "pct_cuda_time": 0.0895162346031788, "trace": "" }, "children": [ @@ -3222,8 +3222,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 27.139, - "pct_cuda_time": 0.09991026520084487, + "cuda_time_us": 24.836, + "pct_cuda_time": 0.0895162346031788, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -3233,9 +3233,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 87.561, - "cuda_time_us": 193.438, - "pct_cuda_time": 0.7121280032396562, + "cpu_time_us": 84.116, + "cuda_time_us": 189.572, + "pct_cuda_time": 0.6832731368253266, "trace": "" }, "children": [ @@ -3243,8 +3243,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 193.438, - "pct_cuda_time": 0.7121280032396562, + "cuda_time_us": 189.572, + "pct_cuda_time": 0.6832731368253266, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -3258,18 +3258,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1039.482, - "cuda_time_us": 875.403, - "pct_cuda_time": 3.222732815785961, + "cpu_time_us": 1145.603, + "cuda_time_us": 859.992, + "pct_cuda_time": 3.099663618491583, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.213, - "cuda_time_us": 15.102, - "pct_cuda_time": 0.05559692048576438, + "cpu_time_us": 31.838, + "cuda_time_us": 14.861, + "pct_cuda_time": 0.053563406443784836, "trace": "" }, "children": [ @@ -3277,8 +3277,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.102, - "pct_cuda_time": 0.05559692048576438, + "cuda_time_us": 14.861, + "pct_cuda_time": 0.053563406443784836, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3288,18 +3288,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 723.42, - "cuda_time_us": 324.455, - "pct_cuda_time": 1.1944576106614142, + "cpu_time_us": 760.807, + "cuda_time_us": 327.722, + "pct_cuda_time": 1.1812062907321212, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.145, - "cuda_time_us": 87.91, - "pct_cuda_time": 0.3236343053836277, + "cpu_time_us": 90.315, + "cuda_time_us": 86.899, + "pct_cuda_time": 0.31320950518528085, "trace": "" }, "children": [ @@ -3307,8 +3307,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 87.91, - "pct_cuda_time": 0.3236343053836277, + "cuda_time_us": 86.899, + "pct_cuda_time": 0.31320950518528085, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -3318,9 +3318,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 74.892, - "cuda_time_us": 36.129, - "pct_cuda_time": 0.13300629984307913, + "cpu_time_us": 87.681, + "cuda_time_us": 36.33, + "pct_cuda_time": 0.13094398466474014, "trace": "" }, "children": [ @@ -3328,8 +3328,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.049, - "pct_cuda_time": 0.02595038355874408, + "cuda_time_us": 7.41, + "pct_cuda_time": 0.026707815204121233, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -3338,8 +3338,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.127, - "pct_cuda_time": 0.018874679600749168, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.017178063058413198, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -3348,8 +3348,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 10.415, - "pct_cuda_time": 0.03834206905437928, + "cuda_time_us": 10.455, + "pct_cuda_time": 0.03768288906330465, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -3358,8 +3358,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.531, - "pct_cuda_time": 0.031406259347374905, + "cuda_time_us": 8.692, + "pct_cuda_time": 0.03132851953498269, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -3369,7 +3369,7 @@ "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 5.007, - "pct_cuda_time": 0.01843290828183169, + "pct_cuda_time": 0.018046697803918356, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -3379,9 +3379,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 430.102, - "cuda_time_us": 136.882, - "pct_cuda_time": 0.5039211806338497, + "cpu_time_us": 448.891, + "cuda_time_us": 137.764, + "pct_cuda_time": 0.4965418965965665, "trace": "" }, "children": [ @@ -3389,8 +3389,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.172, - "pct_cuda_time": 0.03376605447592576, + "cuda_time_us": 9.733, + "pct_cuda_time": 0.035080589120339, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -3399,8 +3399,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 127.71, - "pct_cuda_time": 0.47015512615792393, + "cuda_time_us": 128.031, + "pct_cuda_time": 0.46146130747622754, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -3410,9 +3410,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 98.197, - "cuda_time_us": 63.534, - "pct_cuda_time": 0.23389582480085774, + "cpu_time_us": 98.747, + "cuda_time_us": 66.729, + "pct_cuda_time": 0.24051090428553384, "trace": "" }, "children": [ @@ -3420,8 +3420,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 63.534, - "pct_cuda_time": 0.23389582480085774, + "cuda_time_us": 66.729, + "pct_cuda_time": 0.24051090428553384, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -3433,9 +3433,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 47.601, - "cuda_time_us": 14.22, - "pct_cuda_time": 0.05234990129172093, + "cpu_time_us": 116.775, + "cuda_time_us": 14.501, + "pct_cuda_time": 0.05226586076585182, "trace": "" }, "children": [ @@ -3443,8 +3443,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.22, - "pct_cuda_time": 0.05234990129172093, + "cuda_time_us": 14.501, + "pct_cuda_time": 0.05226586076585182, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3454,18 +3454,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 210.334, - "cuda_time_us": 521.626, - "pct_cuda_time": 1.9203283833470617, + "cpu_time_us": 214.991, + "cuda_time_us": 502.908, + "pct_cuda_time": 1.8126280605498246, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 86.83, - "cuda_time_us": 304.584, - "pct_cuda_time": 1.1213039616763378, + "cpu_time_us": 88.533, + "cuda_time_us": 294.289, + "pct_cuda_time": 1.060703944481192, "trace": "" }, "children": [ @@ -3473,8 +3473,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 304.584, - "pct_cuda_time": 1.1213039616763378, + "cuda_time_us": 294.289, + "pct_cuda_time": 1.060703944481192, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -3484,9 +3484,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.468, - "cuda_time_us": 26.448, - "pct_cuda_time": 0.09736639868941174, + "cpu_time_us": 27.001, + "cuda_time_us": 25.336, + "pct_cuda_time": 0.09131838137808578, "trace": "" }, "children": [ @@ -3494,8 +3494,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 26.448, - "pct_cuda_time": 0.09736639868941174, + "cuda_time_us": 25.336, + "pct_cuda_time": 0.09131838137808578, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -3505,9 +3505,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 80.861, - "cuda_time_us": 190.594, - "pct_cuda_time": 0.7016580229813121, + "cpu_time_us": 81.291, + "cuda_time_us": 183.283, + "pct_cuda_time": 0.6606057346905468, "trace": "" }, "children": [ @@ -3515,8 +3515,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 190.594, - "pct_cuda_time": 0.7016580229813121, + "cuda_time_us": 183.283, + "pct_cuda_time": 0.6606057346905468, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -3530,18 +3530,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1033.744, - "cuda_time_us": 849.0630000000001, - "pct_cuda_time": 3.1257640112835756, + "cpu_time_us": 1116.069, + "cuda_time_us": 855.754, + "pct_cuda_time": 3.0843886224274715, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.8, - "cuda_time_us": 14.3, - "pct_cuda_time": 0.052644415504332576, + "cpu_time_us": 40.11, + "cuda_time_us": 15.382, + "pct_cuda_time": 0.0554412433832379, "trace": "" }, "children": [ @@ -3549,8 +3549,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.3, - "pct_cuda_time": 0.052644415504332576, + "cuda_time_us": 15.382, + "pct_cuda_time": 0.0554412433832379, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3560,18 +3560,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 726.314, - "cuda_time_us": 307.68, - "pct_cuda_time": 1.1327016617044088, + "cpu_time_us": 793.987, + "cuda_time_us": 331.225, + "pct_cuda_time": 1.1938321310371196, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 80.561, - "cuda_time_us": 88.231, - "pct_cuda_time": 0.32481604366173195, + "cpu_time_us": 92.428, + "cuda_time_us": 86.308, + "pct_cuda_time": 0.31107936769734085, "trace": "" }, "children": [ @@ -3579,8 +3579,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 88.231, - "pct_cuda_time": 0.32481604366173195, + "cuda_time_us": 86.308, + "pct_cuda_time": 0.31107936769734085, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -3590,9 +3590,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 74.831, - "cuda_time_us": 36.007999999999996, - "pct_cuda_time": 0.132560847096504, + "cpu_time_us": 78.007, + "cuda_time_us": 35.688, + "pct_cuda_time": 0.1286300282057596, "trace": "" }, "children": [ @@ -3600,8 +3600,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.21, - "pct_cuda_time": 0.026543093411625026, + "cuda_time_us": 7.571, + "pct_cuda_time": 0.027288106465641275, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -3611,7 +3611,7 @@ "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 4.926, - "pct_cuda_time": 0.018134712641562394, + "pct_cuda_time": 0.017754750026383426, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -3620,8 +3620,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 10.615, - "pct_cuda_time": 0.039078354585908415, + "cuda_time_us": 10.415, + "pct_cuda_time": 0.037538717321312094, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -3630,8 +3630,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 9.052, - "pct_cuda_time": 0.033324283157008285, + "cuda_time_us": 8.331, + "pct_cuda_time": 0.030027369563499866, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -3640,8 +3640,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.205, - "pct_cuda_time": 0.015480403300399893, + "cuda_time_us": 4.445, + "pct_cuda_time": 0.016021084828922928, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -3651,9 +3651,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 427.718, - "cuda_time_us": 137.203, - "pct_cuda_time": 0.505102918911954, + "cpu_time_us": 442.091, + "cuda_time_us": 144.133, + "pct_cuda_time": 0.5194976422153315, "trace": "" }, "children": [ @@ -3661,8 +3661,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.453, - "pct_cuda_time": 0.03480053564772418, + "cuda_time_us": 9.934, + "pct_cuda_time": 0.03580505212385159, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -3671,8 +3671,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 127.75, - "pct_cuda_time": 0.4703023832642298, + "cuda_time_us": 134.199, + "pct_cuda_time": 0.4836925900914798, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -3682,9 +3682,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 101.101, - "cuda_time_us": 46.238, - "pct_cuda_time": 0.17022185203421886, + "cpu_time_us": 102.783, + "cuda_time_us": 65.096, + "pct_cuda_time": 0.2346250929186877, "trace": "" }, "children": [ @@ -3692,8 +3692,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 46.238, - "pct_cuda_time": 0.17022185203421886, + "cuda_time_us": 65.096, + "pct_cuda_time": 0.2346250929186877, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -3705,9 +3705,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.328, - "cuda_time_us": 14.18, - "pct_cuda_time": 0.0522026441854151, + "cpu_time_us": 38.878, + "cuda_time_us": 13.9, + "pct_cuda_time": 0.050099680342413654, "trace": "" }, "children": [ @@ -3715,8 +3715,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.18, - "pct_cuda_time": 0.0522026441854151, + "cuda_time_us": 13.9, + "pct_cuda_time": 0.050099680342413654, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3726,18 +3726,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 212.908, - "cuda_time_us": 512.903, - "pct_cuda_time": 1.8882152898894191, + "cpu_time_us": 216.644, + "cuda_time_us": 495.247, + "pct_cuda_time": 1.7850155676647002, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.955, - "cuda_time_us": 299.226, - "pct_cuda_time": 1.1015788722866726, + "cpu_time_us": 85.819, + "cuda_time_us": 290.944, + "pct_cuda_time": 1.0486475825570647, "trace": "" }, "children": [ @@ -3745,8 +3745,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 299.226, - "pct_cuda_time": 1.1015788722866726, + "cuda_time_us": 290.944, + "pct_cuda_time": 1.0486475825570647, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -3756,9 +3756,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 28.202, - "cuda_time_us": 24.526, - "pct_cuda_time": 0.09029069473141682, + "cpu_time_us": 26.62, + "cuda_time_us": 24.986, + "pct_cuda_time": 0.0900568786356509, "trace": "" }, "children": [ @@ -3766,8 +3766,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 24.526, - "pct_cuda_time": 0.09029069473141682, + "cuda_time_us": 24.986, + "pct_cuda_time": 0.0900568786356509, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -3777,9 +3777,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 82.814, - "cuda_time_us": 189.151, - "pct_cuda_time": 0.6963457228713295, + "cpu_time_us": 83.866, + "cuda_time_us": 179.317, + "pct_cuda_time": 0.6463111064719849, "trace": "" }, "children": [ @@ -3787,8 +3787,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 189.151, - "pct_cuda_time": 0.6963457228713295, + "cuda_time_us": 179.317, + "pct_cuda_time": 0.6463111064719849, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -3802,18 +3802,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1046.913, - "cuda_time_us": 840.83, - "pct_cuda_time": 3.095454817378179, + "cpu_time_us": 1094.416, + "cuda_time_us": 827.2729999999999, + "pct_cuda_time": 2.9817347378352204, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 34.561, - "cuda_time_us": 15.262, - "pct_cuda_time": 0.05618594891098768, + "cpu_time_us": 33.86, + "cuda_time_us": 14.301, + "pct_cuda_time": 0.05154500205588904, "trace": "" }, "children": [ @@ -3821,8 +3821,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.262, - "pct_cuda_time": 0.05618594891098768, + "cuda_time_us": 14.301, + "pct_cuda_time": 0.05154500205588904, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3832,18 +3832,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 737.981, - "cuda_time_us": 309.011, - "pct_cuda_time": 1.1376016419167352, + "cpu_time_us": 742.3, + "cuda_time_us": 328.602, + "pct_cuda_time": 1.1843780690559575, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 85.989, - "cuda_time_us": 84.315, - "pct_cuda_time": 0.3103995729543917, + "cpu_time_us": 84.215, + "cuda_time_us": 80.639, + "pct_cuda_time": 0.2906466275634456, "trace": "" }, "children": [ @@ -3851,8 +3851,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 84.315, - "pct_cuda_time": 0.3103995729543917, + "cuda_time_us": 80.639, + "pct_cuda_time": 0.2906466275634456, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -3862,9 +3862,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 79.008, - "cuda_time_us": 36.609, - "pct_cuda_time": 0.13477338511874903, + "cpu_time_us": 78.067, + "cuda_time_us": 38.853, + "pct_cuda_time": 0.1400376172909207, "trace": "" }, "children": [ @@ -3872,8 +3872,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.45, - "pct_cuda_time": 0.02742663604945998, + "cuda_time_us": 7.651, + "pct_cuda_time": 0.02757644994962639, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -3882,8 +3882,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.006, - "pct_cuda_time": 0.018429226854174047, + "cuda_time_us": 5.206, + "pct_cuda_time": 0.01876395222033133, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -3892,8 +3892,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 10.214, - "pct_cuda_time": 0.03760210209519251, + "cuda_time_us": 11.497, + "pct_cuda_time": 0.04143856294221077, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -3902,8 +3902,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.732, - "pct_cuda_time": 0.03214622630656168, + "cuda_time_us": 9.493, + "pct_cuda_time": 0.034215558668383654, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -3912,8 +3912,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.207, - "pct_cuda_time": 0.019169193813360817, + "cuda_time_us": 5.006, + "pct_cuda_time": 0.018043093510368542, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -3923,9 +3923,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 436.602, - "cuda_time_us": 140.40699999999998, - "pct_cuda_time": 0.5168982131270505, + "cpu_time_us": 446.487, + "cuda_time_us": 148.1, + "pct_cuda_time": 0.5337958747274433, "trace": "" }, "children": [ @@ -3933,8 +3933,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.533, - "pct_cuda_time": 0.035095049860335834, + "cuda_time_us": 11.096, + "pct_cuda_time": 0.039993241228735384, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -3943,8 +3943,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 130.874, - "pct_cuda_time": 0.48180316326671474, + "cuda_time_us": 137.004, + "pct_cuda_time": 0.4938026334987079, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -3954,9 +3954,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.97, - "cuda_time_us": 47.68, - "pct_cuda_time": 0.17553047071654385, + "cpu_time_us": 98.527, + "cuda_time_us": 61.01, + "pct_cuda_time": 0.21989794947414798, "trace": "" }, "children": [ @@ -3964,8 +3964,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 47.68, - "pct_cuda_time": 0.17553047071654385, + "cuda_time_us": 61.01, + "pct_cuda_time": 0.21989794947414798, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -3977,9 +3977,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.579, - "cuda_time_us": 14.581, - "pct_cuda_time": 0.053678896676130995, + "cpu_time_us": 72.708, + "cuda_time_us": 14.02, + "pct_cuda_time": 0.050532195568391315, "trace": "" }, "children": [ @@ -3987,8 +3987,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.581, - "pct_cuda_time": 0.053678896676130995, + "cuda_time_us": 14.02, + "pct_cuda_time": 0.050532195568391315, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3998,18 +3998,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 214.34, - "cuda_time_us": 501.976, - "pct_cuda_time": 1.8479883298743252, + "cpu_time_us": 221.411, + "cuda_time_us": 470.34999999999997, + "pct_cuda_time": 1.6952794711549826, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 86.499, - "cuda_time_us": 294.348, - "pct_cuda_time": 1.083620868172677, + "cpu_time_us": 90.004, + "cuda_time_us": 275.431, + "pct_cuda_time": 0.992734176718801, "trace": "" }, "children": [ @@ -4017,8 +4017,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 294.348, - "pct_cuda_time": 1.083620868172677, + "cuda_time_us": 275.431, + "pct_cuda_time": 0.992734176718801, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -4028,9 +4028,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.59, - "cuda_time_us": 25.067, - "pct_cuda_time": 0.09228234709420312, + "cpu_time_us": 29.294, + "cuda_time_us": 25.948, + "pct_cuda_time": 0.0935242090305719, "trace": "" }, "children": [ @@ -4038,8 +4038,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 25.067, - "pct_cuda_time": 0.09228234709420312, + "cuda_time_us": 25.948, + "pct_cuda_time": 0.0935242090305719, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -4049,9 +4049,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 83.725, - "cuda_time_us": 182.561, - "pct_cuda_time": 0.6720851146074447, + "cpu_time_us": 84.125, + "cuda_time_us": 168.971, + "pct_cuda_time": 0.6090210854056098, "trace": "" }, "children": [ @@ -4059,8 +4059,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 182.561, - "pct_cuda_time": 0.6720851146074447, + "cuda_time_us": 168.971, + "pct_cuda_time": 0.6090210854056098, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -4074,18 +4074,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1070.448, - "cuda_time_us": 825.108, - "pct_cuda_time": 3.0375754117446743, + "cpu_time_us": 1133.564, + "cuda_time_us": 813.883, + "pct_cuda_time": 2.9334732472032123, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.083, - "cuda_time_us": 14.741, - "pct_cuda_time": 0.054267925101354295, + "cpu_time_us": 34.672, + "cuda_time_us": 15.062, + "pct_cuda_time": 0.05428786944729744, "trace": "" }, "children": [ @@ -4093,8 +4093,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.741, - "pct_cuda_time": 0.054267925101354295, + "cuda_time_us": 15.062, + "pct_cuda_time": 0.05428786944729744, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4104,18 +4104,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 757.121, - "cuda_time_us": 312.387, - "pct_cuda_time": 1.150030141688947, + "cpu_time_us": 774.498, + "cuda_time_us": 338.437, + "pct_cuda_time": 1.2198262961183777, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 83.905, - "cuda_time_us": 80.079, - "pct_cuda_time": 0.29480504539660474, + "cpu_time_us": 93.039, + "cuda_time_us": 79.618, + "pct_cuda_time": 0.2869666438490856, "trace": "" }, "children": [ @@ -4123,8 +4123,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 80.079, - "pct_cuda_time": 0.29480504539660474, + "cuda_time_us": 79.618, + "pct_cuda_time": 0.2869666438490856, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -4134,9 +4134,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 74.632, - "cuda_time_us": 37.81099999999999, - "pct_cuda_time": 0.13919846116323906, + "cpu_time_us": 77.596, + "cuda_time_us": 38.492000000000004, + "pct_cuda_time": 0.13873646731943787, "trace": "" }, "children": [ @@ -4144,8 +4144,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.69, - "pct_cuda_time": 0.02831017868729493, + "cuda_time_us": 7.811, + "pct_cuda_time": 0.02815313691759662, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -4154,8 +4154,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.367, - "pct_cuda_time": 0.01975822223858412, + "cuda_time_us": 5.206, + "pct_cuda_time": 0.01876395222033133, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -4164,8 +4164,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 10.214, - "pct_cuda_time": 0.03760210209519251, + "cuda_time_us": 11.937, + "pct_cuda_time": 0.0430244521041289, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -4174,8 +4174,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 9.133, - "pct_cuda_time": 0.033622478797277575, + "cuda_time_us": 8.332, + "pct_cuda_time": 0.030030973857049684, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -4184,8 +4184,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.407, - "pct_cuda_time": 0.019905479344889947, + "cuda_time_us": 5.206, + "pct_cuda_time": 0.01876395222033133, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -4195,9 +4195,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 458.775, - "cuda_time_us": 146.697, - "pct_cuda_time": 0.5400543930936417, + "cpu_time_us": 460.509, + "cuda_time_us": 151.785, + "pct_cuda_time": 0.5470776964585076, "trace": "" }, "children": [ @@ -4205,8 +4205,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 10.054, - "pct_cuda_time": 0.03701307366996921, + "cuda_time_us": 9.814, + "pct_cuda_time": 0.035372536897873924, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -4215,8 +4215,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 136.643, - "pct_cuda_time": 0.5030413194236724, + "cuda_time_us": 141.971, + "pct_cuda_time": 0.5117051595606337, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -4226,9 +4226,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 103.434, - "cuda_time_us": 47.8, - "pct_cuda_time": 0.1759722420354613, + "cpu_time_us": 106.609, + "cuda_time_us": 68.542, + "pct_cuda_time": 0.2470454884913465, "trace": "" }, "children": [ @@ -4236,8 +4236,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 47.8, - "pct_cuda_time": 0.1759722420354613, + "cuda_time_us": 68.542, + "pct_cuda_time": 0.2470454884913465, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -4249,9 +4249,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.339, - "cuda_time_us": 15.102, - "pct_cuda_time": 0.05559692048576438, + "cpu_time_us": 55.343, + "cuda_time_us": 15.422, + "pct_cuda_time": 0.05558541512523046, "trace": "" }, "children": [ @@ -4259,8 +4259,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.102, - "pct_cuda_time": 0.05559692048576438, + "cuda_time_us": 15.422, + "pct_cuda_time": 0.05558541512523046, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4270,18 +4270,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 217.745, - "cuda_time_us": 482.878, - "pct_cuda_time": 1.7776804244686089, + "cpu_time_us": 246.298, + "cuda_time_us": 444.962, + "pct_cuda_time": 1.6037736665123068, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 86.669, - "cuda_time_us": 280.408, - "pct_cuda_time": 1.032301766625097, + "cpu_time_us": 107.561, + "cuda_time_us": 258.395, + "pct_cuda_time": 0.9313314318041709, "trace": "" }, "children": [ @@ -4289,8 +4289,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 280.408, - "pct_cuda_time": 1.032301766625097, + "cuda_time_us": 258.395, + "pct_cuda_time": 0.9313314318041709, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -4300,9 +4300,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 28.953, - "cuda_time_us": 25.436, - "pct_cuda_time": 0.09364079389987436, + "cpu_time_us": 32.268, + "cuda_time_us": 26.258, + "pct_cuda_time": 0.09464154003101422, "trace": "" }, "children": [ @@ -4310,8 +4310,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 25.436, - "pct_cuda_time": 0.09364079389987436, + "cuda_time_us": 26.258, + "pct_cuda_time": 0.09464154003101422, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -4321,9 +4321,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 84.847, - "cuda_time_us": 177.034, - "pct_cuda_time": 0.6517378639436373, + "cpu_time_us": 87.541, + "cuda_time_us": 160.309, + "pct_cuda_time": 0.5778006946771216, "trace": "" }, "children": [ @@ -4331,8 +4331,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 177.034, - "pct_cuda_time": 0.6517378639436373, + "cuda_time_us": 160.309, + "pct_cuda_time": 0.5778006946771216, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -4346,18 +4346,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1050.288, - "cuda_time_us": 812.1390000000001, - "pct_cuda_time": 2.9898309764526685, + "cpu_time_us": 1059.514, + "cuda_time_us": 785.601, + "pct_cuda_time": 2.8315366170273744, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.773, - "cuda_time_us": 15.342, - "pct_cuda_time": 0.05648046312359933, + "cpu_time_us": 34.672, + "cuda_time_us": 14.781, + "pct_cuda_time": 0.053275062959799724, "trace": "" }, "children": [ @@ -4365,8 +4365,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.342, - "pct_cuda_time": 0.05648046312359933, + "cuda_time_us": 14.781, + "pct_cuda_time": 0.053275062959799724, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4376,18 +4376,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 740.075, - "cuda_time_us": 314.641, - "pct_cuda_time": 1.1583280796292803, + "cpu_time_us": 742.46, + "cuda_time_us": 335.634, + "pct_cuda_time": 1.209723461298249, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.295, - "cuda_time_us": 79.938, - "pct_cuda_time": 0.29428596409687674, + "cpu_time_us": 82.654, + "cuda_time_us": 72.127, + "pct_cuda_time": 0.2599668808674295, "trace": "" }, "children": [ @@ -4395,8 +4395,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 79.938, - "pct_cuda_time": 0.29428596409687674, + "cuda_time_us": 72.127, + "pct_cuda_time": 0.2599668808674295, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -4406,9 +4406,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 73.89, - "cuda_time_us": 38.812999999999995, - "pct_cuda_time": 0.1428872516762, + "cpu_time_us": 75.933, + "cuda_time_us": 38.693000000000005, + "pct_cuda_time": 0.13946093032295048, "trace": "" }, "children": [ @@ -4416,8 +4416,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.731, - "pct_cuda_time": 0.028461117221258398, + "cuda_time_us": 8.011, + "pct_cuda_time": 0.028873995627559403, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -4426,8 +4426,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.246, - "pct_cuda_time": 0.019312769492009, + "cuda_time_us": 5.527, + "pct_cuda_time": 0.0199209304498216, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -4436,8 +4436,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.577, - "pct_cuda_time": 0.04261988799256351, + "cuda_time_us": 11.096, + "pct_cuda_time": 0.039993241228735384, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -4446,8 +4446,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.892, - "pct_cuda_time": 0.03273525473178498, + "cuda_time_us": 8.852, + "pct_cuda_time": 0.03190520650295292, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -4456,8 +4456,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.367, - "pct_cuda_time": 0.01975822223858412, + "cuda_time_us": 5.207, + "pct_cuda_time": 0.018767556513881144, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -4467,9 +4467,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 445.224, - "cuda_time_us": 149.982, - "pct_cuda_time": 0.5521478829490075, + "cpu_time_us": 447.489, + "cuda_time_us": 155.551, + "pct_cuda_time": 0.5606514659671069, "trace": "" }, "children": [ @@ -4477,8 +4477,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 10.415, - "pct_cuda_time": 0.03834206905437928, + "cuda_time_us": 10.214, + "pct_cuda_time": 0.0368142543177995, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -4487,8 +4487,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 139.567, - "pct_cuda_time": 0.5138058138946283, + "cuda_time_us": 145.337, + "pct_cuda_time": 0.5238372116493074, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -4498,9 +4498,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 101.342, - "cuda_time_us": 45.908, - "pct_cuda_time": 0.1690069809071958, + "cpu_time_us": 100.55, + "cuda_time_us": 69.263, + "pct_cuda_time": 0.24964418414076237, "trace": "" }, "children": [ @@ -4508,8 +4508,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 45.908, - "pct_cuda_time": 0.1690069809071958, + "cuda_time_us": 69.263, + "pct_cuda_time": 0.24964418414076237, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -4521,9 +4521,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.409, - "cuda_time_us": 14.701, - "pct_cuda_time": 0.05412066799504847, + "cpu_time_us": 38.868, + "cuda_time_us": 15.622, + "pct_cuda_time": 0.05630627383519324, "trace": "" }, "children": [ @@ -4531,8 +4531,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.701, - "pct_cuda_time": 0.05412066799504847, + "cuda_time_us": 15.622, + "pct_cuda_time": 0.05630627383519324, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4542,18 +4542,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 214.861, - "cuda_time_us": 467.45500000000004, - "pct_cuda_time": 1.7209017657047403, + "cpu_time_us": 221.852, + "cuda_time_us": 419.56399999999996, + "pct_cuda_time": 1.5122318189341324, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 86.86, - "cuda_time_us": 273.687, - "pct_cuda_time": 1.0075588913380609, + "cpu_time_us": 85.378, + "cuda_time_us": 250.243, + "pct_cuda_time": 0.9019492307860876, "trace": "" }, "children": [ @@ -4561,8 +4561,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 273.687, - "pct_cuda_time": 1.0075588913380609, + "cuda_time_us": 250.243, + "pct_cuda_time": 0.9019492307860876, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -4572,9 +4572,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.008, - "cuda_time_us": 25.768, - "pct_cuda_time": 0.09486302788221272, + "cpu_time_us": 25.849, + "cuda_time_us": 23.755, + "pct_cuda_time": 0.08561999327582993, "trace": "" }, "children": [ @@ -4582,8 +4582,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 25.768, - "pct_cuda_time": 0.09486302788221272, + "cuda_time_us": 23.755, + "pct_cuda_time": 0.08561999327582993, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -4593,9 +4593,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 84.186, - "cuda_time_us": 168.0, - "pct_cuda_time": 0.6184798464844666, + "cpu_time_us": 92.879, + "cuda_time_us": 145.566, + "pct_cuda_time": 0.5246625948722148, "trace": "" }, "children": [ @@ -4603,8 +4603,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 168.0, - "pct_cuda_time": 0.6184798464844666, + "cuda_time_us": 145.566, + "pct_cuda_time": 0.5246625948722148, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -4618,18 +4618,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1038.64, - "cuda_time_us": 819.1789999999999, - "pct_cuda_time": 3.0157482271624927, + "cpu_time_us": 1069.319, + "cuda_time_us": 784.348, + "pct_cuda_time": 2.8270204372094576, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 36.064, - "cuda_time_us": 15.301, - "pct_cuda_time": 0.05632952458963585, + "cpu_time_us": 35.062, + "cuda_time_us": 15.022, + "pct_cuda_time": 0.054143697705304875, "trace": "" }, "children": [ @@ -4637,8 +4637,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.301, - "pct_cuda_time": 0.05632952458963585, + "cuda_time_us": 15.022, + "pct_cuda_time": 0.054143697705304875, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4648,18 +4648,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 735.178, - "cuda_time_us": 342.90399999999994, - "pct_cuda_time": 1.2623762695173184, + "cpu_time_us": 755.73, + "cuda_time_us": 336.93399999999997, + "pct_cuda_time": 1.2144090429130072, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 82.323, - "cuda_time_us": 68.021, - "pct_cuda_time": 0.25041439070071375, + "cpu_time_us": 83.675, + "cuda_time_us": 67.62, + "pct_cuda_time": 0.24372232983841807, "trace": "" }, "children": [ @@ -4667,8 +4667,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 68.021, - "pct_cuda_time": 0.25041439070071375, + "cuda_time_us": 67.62, + "pct_cuda_time": 0.24372232983841807, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -4678,9 +4678,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 76.475, - "cuda_time_us": 40.656, - "pct_cuda_time": 0.14967212284924092, + "cpu_time_us": 74.662, + "cuda_time_us": 39.614000000000004, + "pct_cuda_time": 0.14278048468232912, "trace": "" }, "children": [ @@ -4688,8 +4688,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.851, - "pct_cuda_time": 0.028902888540175878, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.029306510853537082, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -4698,8 +4698,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.166, - "pct_cuda_time": 0.01901825527939735, + "cuda_time_us": 5.647, + "pct_cuda_time": 0.020353445675799272, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -4708,8 +4708,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 12.939, - "pct_cuda_time": 0.04763399246227686, + "cuda_time_us": 11.657, + "pct_cuda_time": 0.042015249910181005, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -4718,8 +4718,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 9.172, - "pct_cuda_time": 0.03376605447592576, + "cuda_time_us": 8.972, + "pct_cuda_time": 0.03233772172893059, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -4728,8 +4728,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.528, - "pct_cuda_time": 0.020350932091465067, + "cuda_time_us": 5.207, + "pct_cuda_time": 0.018767556513881144, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -4739,9 +4739,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 439.546, - "cuda_time_us": 154.38799999999998, - "pct_cuda_time": 0.5683682532085941, + "cpu_time_us": 460.919, + "cuda_time_us": 157.954, + "pct_cuda_time": 0.5693125833673098, "trace": "" }, "children": [ @@ -4749,8 +4749,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 10.414, - "pct_cuda_time": 0.038338387626721634, + "cuda_time_us": 10.334, + "pct_cuda_time": 0.03724676954377717, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -4759,8 +4759,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 143.974, - "pct_cuda_time": 0.5300298655818725, + "cuda_time_us": 147.62, + "pct_cuda_time": 0.5320658138235326, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -4770,9 +4770,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 100.851, - "cuda_time_us": 79.839, - "pct_cuda_time": 0.2939215027587698, + "cpu_time_us": 101.021, + "cuda_time_us": 71.746, + "pct_cuda_time": 0.2585936450249503, "trace": "" }, "children": [ @@ -4780,8 +4780,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 79.839, - "pct_cuda_time": 0.2939215027587698, + "cuda_time_us": 71.746, + "pct_cuda_time": 0.2585936450249503, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -4793,9 +4793,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.536, - "cuda_time_us": 15.141, - "pct_cuda_time": 0.055740496164412554, + "cpu_time_us": 40.14, + "cuda_time_us": 15.062, + "pct_cuda_time": 0.05428786944729744, "trace": "" }, "children": [ @@ -4803,8 +4803,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.141, - "pct_cuda_time": 0.055740496164412554, + "cuda_time_us": 15.062, + "pct_cuda_time": 0.05428786944729744, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4814,18 +4814,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 209.103, - "cuda_time_us": 445.83299999999997, - "pct_cuda_time": 1.641301936891126, + "cpu_time_us": 215.732, + "cuda_time_us": 417.33000000000004, + "pct_cuda_time": 1.5041798271438482, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.264, - "cuda_time_us": 271.424, - "pct_cuda_time": 0.9992278205488087, + "cpu_time_us": 88.563, + "cuda_time_us": 253.788, + "pct_cuda_time": 0.914726451420178, "trace": "" }, "children": [ @@ -4833,8 +4833,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 271.424, - "pct_cuda_time": 0.9992278205488087, + "cuda_time_us": 253.788, + "pct_cuda_time": 0.914726451420178, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -4844,9 +4844,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.61, - "cuda_time_us": 26.679, - "pct_cuda_time": 0.09821680847832788, + "cpu_time_us": 26.6, + "cuda_time_us": 23.714, + "pct_cuda_time": 0.08547221724028757, "trace": "" }, "children": [ @@ -4854,8 +4854,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 26.679, - "pct_cuda_time": 0.09821680847832788, + "cuda_time_us": 23.714, + "pct_cuda_time": 0.08547221724028757, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -4865,9 +4865,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 82.093, - "cuda_time_us": 147.73, - "pct_cuda_time": 0.5438573078639896, + "cpu_time_us": 82.222, + "cuda_time_us": 139.828, + "pct_cuda_time": 0.5039811584833824, "trace": "" }, "children": [ @@ -4875,8 +4875,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 147.73, - "pct_cuda_time": 0.5438573078639896, + "cuda_time_us": 139.828, + "pct_cuda_time": 0.5039811584833824, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -4890,18 +4890,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1021.726, - "cuda_time_us": 760.762, - "pct_cuda_time": 2.800690267685808, + "cpu_time_us": 1095.287, + "cuda_time_us": 750.648, + "pct_cuda_time": 2.7055557445807286, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.139, - "cuda_time_us": 15.222, - "pct_cuda_time": 0.056038691804681844, + "cpu_time_us": 42.884, + "cuda_time_us": 15.102, + "pct_cuda_time": 0.054432041189289994, "trace": "" }, "children": [ @@ -4909,8 +4909,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.222, - "pct_cuda_time": 0.056038691804681844, + "cuda_time_us": 15.102, + "pct_cuda_time": 0.054432041189289994, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4920,18 +4920,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 720.145, - "cuda_time_us": 312.05699999999996, - "pct_cuda_time": 1.1488152705619237, + "cpu_time_us": 771.353, + "cuda_time_us": 339.298, + "pct_cuda_time": 1.2229295928647674, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 80.55, - "cuda_time_us": 66.217, - "pct_cuda_time": 0.24377309520632098, + "cpu_time_us": 87.29, + "cuda_time_us": 67.179, + "pct_cuda_time": 0.2421328363829501, "trace": "" }, "children": [ @@ -4939,8 +4939,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 66.217, - "pct_cuda_time": 0.24377309520632098, + "cuda_time_us": 67.179, + "pct_cuda_time": 0.2421328363829501, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -4950,9 +4950,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 74.18, - "cuda_time_us": 39.895, - "pct_cuda_time": 0.1468705564017726, + "cpu_time_us": 74.822, + "cuda_time_us": 41.337, + "pct_cuda_time": 0.14899068246865851, "trace": "" }, "children": [ @@ -4960,8 +4960,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.212, - "pct_cuda_time": 0.030231883924585947, + "cuda_time_us": 8.612, + "pct_cuda_time": 0.031040176050997584, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -4970,8 +4970,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.527, - "pct_cuda_time": 0.020347250663807423, + "cuda_time_us": 5.568, + "pct_cuda_time": 0.02006870648536397, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -4980,8 +4980,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.776, - "pct_cuda_time": 0.04335249209643499, + "cuda_time_us": 12.417, + "pct_cuda_time": 0.044754513008039586, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -4990,8 +4990,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.813, - "pct_cuda_time": 0.03244442194683098, + "cuda_time_us": 9.413, + "pct_cuda_time": 0.03392721518439854, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -5000,8 +5000,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.567, - "pct_cuda_time": 0.020494507770113246, + "cuda_time_us": 5.327, + "pct_cuda_time": 0.019200071739858812, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -5011,9 +5011,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 431.153, - "cuda_time_us": 158.875, - "pct_cuda_time": 0.5848868191084502, + "cpu_time_us": 471.174, + "cuda_time_us": 165.525, + "pct_cuda_time": 0.5966006898329511, "trace": "" }, "children": [ @@ -5021,8 +5021,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.814, - "pct_cuda_time": 0.03612953103213425, + "cuda_time_us": 10.054, + "pct_cuda_time": 0.03623756734982927, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -5031,8 +5031,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 149.061, - "pct_cuda_time": 0.548757288076316, + "cuda_time_us": 155.471, + "pct_cuda_time": 0.5603631224831218, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -5042,9 +5042,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 98.537, - "cuda_time_us": 47.07, - "pct_cuda_time": 0.17328479984538, + "cpu_time_us": 102.172, + "cuda_time_us": 65.257, + "pct_cuda_time": 0.23520538418020775, "trace": "" }, "children": [ @@ -5052,8 +5052,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 47.07, - "pct_cuda_time": 0.17328479984538, + "cuda_time_us": 65.257, + "pct_cuda_time": 0.23520538418020775, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -5065,9 +5065,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.767, - "cuda_time_us": 14.981, - "pct_cuda_time": 0.05515146773918925, + "cpu_time_us": 40.341, + "cuda_time_us": 15.702, + "pct_cuda_time": 0.05659461731917835, "trace": "" }, "children": [ @@ -5075,8 +5075,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.981, - "pct_cuda_time": 0.05515146773918925, + "cuda_time_us": 15.702, + "pct_cuda_time": 0.05659461731917835, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5086,18 +5086,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 210.484, - "cuda_time_us": 418.502, - "pct_cuda_time": 1.5406848375800133, + "cpu_time_us": 215.862, + "cuda_time_us": 380.54600000000005, + "pct_cuda_time": 1.3715994932074926, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.856, - "cuda_time_us": 248.059, - "pct_cuda_time": 0.9132112633279184, + "cpu_time_us": 88.342, + "cuda_time_us": 217.785, + "pct_cuda_time": 0.7849610707462271, "trace": "" }, "children": [ @@ -5105,8 +5105,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 248.059, - "pct_cuda_time": 0.9132112633279184, + "cuda_time_us": 217.785, + "pct_cuda_time": 0.7849610707462271, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -5116,9 +5116,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.219, - "cuda_time_us": 25.277, - "pct_cuda_time": 0.09305544690230871, + "cpu_time_us": 26.409, + "cuda_time_us": 24.515, + "pct_cuda_time": 0.08835925637368854, "trace": "" }, "children": [ @@ -5126,8 +5126,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 25.277, - "pct_cuda_time": 0.09305544690230871, + "cuda_time_us": 24.515, + "pct_cuda_time": 0.08835925637368854, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -5137,9 +5137,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 82.102, - "cuda_time_us": 145.166, - "pct_cuda_time": 0.5344181273497861, + "cpu_time_us": 82.593, + "cuda_time_us": 138.246, + "pct_cuda_time": 0.49827916608757683, "trace": "" }, "children": [ @@ -5147,8 +5147,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 145.166, - "pct_cuda_time": 0.5344181273497861, + "cuda_time_us": 138.246, + "pct_cuda_time": 0.49827916608757683, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -5162,18 +5162,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1054.945, - "cuda_time_us": 752.4279999999999, - "pct_cuda_time": 2.770009249586989, + "cpu_time_us": 1166.604, + "cuda_time_us": 766.471, + "pct_cuda_time": 2.762586481419434, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.208, - "cuda_time_us": 16.383, - "pct_cuda_time": 0.06031282931520843, + "cpu_time_us": 32.368, + "cuda_time_us": 16.183, + "pct_cuda_time": 0.05832828251663885, "trace": "" }, "children": [ @@ -5181,8 +5181,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.383, - "pct_cuda_time": 0.06031282931520843, + "cuda_time_us": 16.183, + "pct_cuda_time": 0.05832828251663885, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5192,18 +5192,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 746.154, - "cuda_time_us": 310.734, - "pct_cuda_time": 1.1439447417708586, + "cpu_time_us": 854.607, + "cuda_time_us": 345.907, + "pct_cuda_time": 1.2467503689354877, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 80.671, - "cuda_time_us": 66.018, - "pct_cuda_time": 0.24304049110244952, + "cpu_time_us": 89.343, + "cuda_time_us": 66.899, + "pct_cuda_time": 0.2411236341890022, "trace": "" }, "children": [ @@ -5211,8 +5211,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 66.018, - "pct_cuda_time": 0.24304049110244952, + "cuda_time_us": 66.899, + "pct_cuda_time": 0.2411236341890022, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -5222,9 +5222,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 73.33, - "cuda_time_us": 40.214999999999996, - "pct_cuda_time": 0.14804861325221919, + "cpu_time_us": 77.926, + "cuda_time_us": 41.17699999999999, + "pct_cuda_time": 0.14841399550068823, "trace": "" }, "children": [ @@ -5232,8 +5232,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.029786431178010823, + "cuda_time_us": 9.173, + "pct_cuda_time": 0.03306218473244319, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -5242,8 +5242,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.487, - "pct_cuda_time": 0.020199993557501596, + "cuda_time_us": 5.327, + "pct_cuda_time": 0.019200071739858812, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -5252,8 +5252,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.777, - "pct_cuda_time": 0.043356173524092635, + "cuda_time_us": 11.897, + "pct_cuda_time": 0.04288028036213635, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -5262,8 +5262,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 9.252, - "pct_cuda_time": 0.034060568688537414, + "cuda_time_us": 9.373, + "pct_cuda_time": 0.03378304344240598, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -5272,8 +5272,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.608, - "pct_cuda_time": 0.020645446304076717, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.019488415223843928, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -5283,9 +5283,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 447.218, - "cuda_time_us": 158.714, - "pct_cuda_time": 0.5842941092555692, + "cpu_time_us": 464.535, + "cuda_time_us": 165.244, + "pct_cuda_time": 0.5955878833454533, "trace": "" }, "children": [ @@ -5294,7 +5294,7 @@ "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 10.294, - "pct_cuda_time": 0.037896616307804165, + "pct_cuda_time": 0.03710259780178462, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -5303,8 +5303,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 148.42, - "pct_cuda_time": 0.546397492947765, + "cuda_time_us": 154.95, + "pct_cuda_time": 0.5584852855436687, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -5314,9 +5314,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 101.271, - "cuda_time_us": 45.787, - "pct_cuda_time": 0.16856152816062064, + "cpu_time_us": 101.381, + "cuda_time_us": 72.587, + "pct_cuda_time": 0.2616248559003439, "trace": "" }, "children": [ @@ -5324,8 +5324,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 45.787, - "pct_cuda_time": 0.16856152816062064, + "cuda_time_us": 72.587, + "pct_cuda_time": 0.2616248559003439, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -5337,9 +5337,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.796, - "cuda_time_us": 14.902, - "pct_cuda_time": 0.05486063495423524, + "cpu_time_us": 37.416, + "cuda_time_us": 16.184, + "pct_cuda_time": 0.05833188681018867, "trace": "" }, "children": [ @@ -5347,8 +5347,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.902, - "pct_cuda_time": 0.05486063495423524, + "cuda_time_us": 16.184, + "pct_cuda_time": 0.05833188681018867, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5358,18 +5358,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 218.977, - "cuda_time_us": 410.409, - "pct_cuda_time": 1.5108910435466871, + "cpu_time_us": 220.9, + "cuda_time_us": 388.197, + "pct_cuda_time": 1.3991759431571187, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 86.71, - "cuda_time_us": 250.092, - "pct_cuda_time": 0.9206956057559121, + "cpu_time_us": 92.759, + "cuda_time_us": 225.035, + "pct_cuda_time": 0.811092198982378, "trace": "" }, "children": [ @@ -5377,8 +5377,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 250.092, - "pct_cuda_time": 0.9206956057559121, + "cuda_time_us": 225.035, + "pct_cuda_time": 0.811092198982378, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -5388,9 +5388,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.37, - "cuda_time_us": 23.434, - "pct_cuda_time": 0.0862705757292678, + "cpu_time_us": 26.85, + "cuda_time_us": 24.436, + "pct_cuda_time": 0.08807451718325324, "trace": "" }, "children": [ @@ -5398,8 +5398,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 23.434, - "pct_cuda_time": 0.0862705757292678, + "cuda_time_us": 24.436, + "pct_cuda_time": 0.08807451718325324, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -5409,9 +5409,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 87.381, - "cuda_time_us": 136.883, - "pct_cuda_time": 0.5039248620615074, + "cpu_time_us": 83.505, + "cuda_time_us": 138.726, + "pct_cuda_time": 0.5000092269914874, "trace": "" }, "children": [ @@ -5419,8 +5419,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 136.883, - "pct_cuda_time": 0.5039248620615074, + "cuda_time_us": 138.726, + "pct_cuda_time": 0.5000092269914874, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -5434,18 +5434,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1031.66, - "cuda_time_us": 743.9680000000001, - "pct_cuda_time": 2.7388643716033076, + "cpu_time_us": 1049.97, + "cuda_time_us": 777.477, + "pct_cuda_time": 2.8022553362286864, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 34.872, - "cuda_time_us": 16.504, - "pct_cuda_time": 0.06075828206178356, + "cpu_time_us": 33.42, + "cuda_time_us": 16.223, + "pct_cuda_time": 0.05847245425863141, "trace": "" }, "children": [ @@ -5453,8 +5453,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.504, - "pct_cuda_time": 0.06075828206178356, + "cuda_time_us": 16.223, + "pct_cuda_time": 0.05847245425863141, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5464,18 +5464,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 725.443, - "cuda_time_us": 316.024, - "pct_cuda_time": 1.1634194940798042, + "cpu_time_us": 743.701, + "cuda_time_us": 336.974, + "pct_cuda_time": 1.2145532146549998, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 82.232, - "cuda_time_us": 67.179, - "pct_cuda_time": 0.24731462861297607, + "cpu_time_us": 80.47, + "cuda_time_us": 67.42, + "pct_cuda_time": 0.24300147112845527, "trace": "" }, "children": [ @@ -5483,8 +5483,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 67.179, - "pct_cuda_time": 0.24731462861297607, + "cuda_time_us": 67.42, + "pct_cuda_time": 0.24300147112845527, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -5494,9 +5494,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 75.643, - "cuda_time_us": 40.295, - "pct_cuda_time": 0.14834312746483086, + "cpu_time_us": 75.763, + "cuda_time_us": 40.415, + "pct_cuda_time": 0.14566752381573006, "trace": "" }, "children": [ @@ -5504,8 +5504,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.372, - "pct_cuda_time": 0.03082091234980925, + "cuda_time_us": 8.331, + "pct_cuda_time": 0.030027369563499866, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -5514,8 +5514,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.407, - "pct_cuda_time": 0.019905479344889947, + "cuda_time_us": 5.488, + "pct_cuda_time": 0.019780363001378858, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -5524,8 +5524,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.496, - "pct_cuda_time": 0.042321692352294216, + "cuda_time_us": 12.257, + "pct_cuda_time": 0.04417782604006936, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -5534,8 +5534,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 9.413, - "pct_cuda_time": 0.03465327854141836, + "cuda_time_us": 9.012, + "pct_cuda_time": 0.03248189347092315, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -5544,8 +5544,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.607, - "pct_cuda_time": 0.020641764876419073, + "cuda_time_us": 5.327, + "pct_cuda_time": 0.019200071739858812, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -5555,9 +5555,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 432.205, - "cuda_time_us": 161.48, - "pct_cuda_time": 0.5944769381566171, + "cpu_time_us": 444.535, + "cuda_time_us": 164.76399999999998, + "pct_cuda_time": 0.5938578224415425, "trace": "" }, "children": [ @@ -5565,8 +5565,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 10.856, - "pct_cuda_time": 0.039965578651401004, + "cuda_time_us": 10.695, + "pct_cuda_time": 0.03854791951526, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -5575,8 +5575,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 150.624, - "pct_cuda_time": 0.554511359505216, + "cuda_time_us": 154.069, + "pct_cuda_time": 0.5553099029262826, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -5586,9 +5586,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 100.049, - "cuda_time_us": 47.07, - "pct_cuda_time": 0.17328479984538, + "cpu_time_us": 107.371, + "cuda_time_us": 64.375, + "pct_cuda_time": 0.23202639726927185, "trace": "" }, "children": [ @@ -5596,8 +5596,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 47.07, - "pct_cuda_time": 0.17328479984538, + "cuda_time_us": 64.375, + "pct_cuda_time": 0.23202639726927185, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -5609,9 +5609,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.127, - "cuda_time_us": 14.3, - "pct_cuda_time": 0.052644415504332576, + "cpu_time_us": 38.117, + "cuda_time_us": 15.903, + "pct_cuda_time": 0.05731908032269096, "trace": "" }, "children": [ @@ -5619,8 +5619,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.3, - "pct_cuda_time": 0.052644415504332576, + "cuda_time_us": 15.903, + "pct_cuda_time": 0.05731908032269096, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5630,18 +5630,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 211.205, - "cuda_time_us": 397.14, - "pct_cuda_time": 1.4620421799573873, + "cpu_time_us": 212.928, + "cuda_time_us": 408.37699999999995, + "pct_cuda_time": 1.4719105869923639, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 84.245, - "cuda_time_us": 233.698, - "pct_cuda_time": 0.8603422807364696, + "cpu_time_us": 84.527, + "cuda_time_us": 239.707, + "pct_cuda_time": 0.8639743939452481, "trace": "" }, "children": [ @@ -5649,8 +5649,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 233.698, - "pct_cuda_time": 0.8603422807364696, + "cuda_time_us": 239.707, + "pct_cuda_time": 0.8639743939452481, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -5660,9 +5660,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.919, - "cuda_time_us": 23.754, - "pct_cuda_time": 0.08744863257971441, + "cpu_time_us": 27.741, + "cuda_time_us": 24.916, + "pct_cuda_time": 0.08980457808716392, "trace": "" }, "children": [ @@ -5670,8 +5670,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 23.754, - "pct_cuda_time": 0.08744863257971441, + "cuda_time_us": 24.916, + "pct_cuda_time": 0.08980457808716392, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -5681,9 +5681,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 82.434, - "cuda_time_us": 139.688, - "pct_cuda_time": 0.5142512666412034, + "cpu_time_us": 82.433, + "cuda_time_us": 143.754, + "pct_cuda_time": 0.518131614959952, "trace": "" }, "children": [ @@ -5691,8 +5691,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 139.688, - "pct_cuda_time": 0.5142512666412034, + "cuda_time_us": 143.754, + "pct_cuda_time": 0.518131614959952, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -5706,18 +5706,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1026.042, - "cuda_time_us": 740.802, - "pct_cuda_time": 2.7272089716392016, + "cpu_time_us": 1080.054, + "cuda_time_us": 799.452, + "pct_cuda_time": 2.8814596869858473, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 34.211, - "cuda_time_us": 15.141, - "pct_cuda_time": 0.055740496164412554, + "cpu_time_us": 34.461, + "cuda_time_us": 16.264, + "pct_cuda_time": 0.058620230294173785, "trace": "" }, "children": [ @@ -5725,8 +5725,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.141, - "pct_cuda_time": 0.055740496164412554, + "cuda_time_us": 16.264, + "pct_cuda_time": 0.058620230294173785, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5736,18 +5736,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 719.184, - "cuda_time_us": 317.466, - "pct_cuda_time": 1.1687281127621292, + "cpu_time_us": 757.252, + "cuda_time_us": 335.89300000000003, + "pct_cuda_time": 1.210656973327651, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 79.859, - "cuda_time_us": 66.539, - "pct_cuda_time": 0.24495851491208287, + "cpu_time_us": 87.051, + "cuda_time_us": 68.141, + "pct_cuda_time": 0.24560016677787114, "trace": "" }, "children": [ @@ -5755,8 +5755,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 66.539, - "pct_cuda_time": 0.24495851491208287, + "cuda_time_us": 68.141, + "pct_cuda_time": 0.24560016677787114, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -5766,9 +5766,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 73.871, - "cuda_time_us": 40.535, - "pct_cuda_time": 0.1492266701026658, + "cpu_time_us": 76.244, + "cuda_time_us": 40.175000000000004, + "pct_cuda_time": 0.1448024933637747, "trace": "" }, "children": [ @@ -5776,8 +5776,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.411, - "pct_cuda_time": 0.030964488028457433, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.029306510853537082, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -5786,8 +5786,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.448, - "pct_cuda_time": 0.02005641787885342, + "cuda_time_us": 5.567, + "pct_cuda_time": 0.020065102191814156, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -5796,8 +5796,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.936, - "pct_cuda_time": 0.0439415205216583, + "cuda_time_us": 11.576, + "pct_cuda_time": 0.04172330213264608, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -5806,8 +5806,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 9.013, - "pct_cuda_time": 0.0331807074783601, + "cuda_time_us": 9.453, + "pct_cuda_time": 0.03407138692639109, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -5816,8 +5816,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.727, - "pct_cuda_time": 0.021083536195336552, + "cuda_time_us": 5.448, + "pct_cuda_time": 0.0196361912593863, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -5827,9 +5827,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 433.277, - "cuda_time_us": 164.44400000000002, - "pct_cuda_time": 0.6053886897338788, + "cpu_time_us": 447.88, + "cuda_time_us": 162.24, + "pct_cuda_time": 0.5847605855218123, "trace": "" }, "children": [ @@ -5837,8 +5837,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.894, - "pct_cuda_time": 0.03642404524474591, + "cuda_time_us": 10.454, + "pct_cuda_time": 0.037679284769754844, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -5847,8 +5847,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 154.55, - "pct_cuda_time": 0.5689646444891329, + "cuda_time_us": 151.786, + "pct_cuda_time": 0.5470813007520574, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -5858,9 +5858,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 97.295, - "cuda_time_us": 45.948, - "pct_cuda_time": 0.16915423801350163, + "cpu_time_us": 109.834, + "cuda_time_us": 65.337, + "pct_cuda_time": 0.23549372766419285, "trace": "" }, "children": [ @@ -5868,8 +5868,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 45.948, - "pct_cuda_time": 0.16915423801350163, + "cuda_time_us": 65.337, + "pct_cuda_time": 0.23549372766419285, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -5881,9 +5881,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.217, - "cuda_time_us": 15.862, - "pct_cuda_time": 0.058394805505575055, + "cpu_time_us": 38.027, + "cuda_time_us": 15.462, + "pct_cuda_time": 0.05572958686722301, "trace": "" }, "children": [ @@ -5891,8 +5891,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.862, - "pct_cuda_time": 0.058394805505575055, + "cuda_time_us": 15.462, + "pct_cuda_time": 0.05572958686722301, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5902,18 +5902,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 213.128, - "cuda_time_us": 392.33299999999997, - "pct_cuda_time": 1.4443455572070847, + "cpu_time_us": 220.74, + "cuda_time_us": 431.83299999999997, + "pct_cuda_time": 1.5564528964967994, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 86.259, - "cuda_time_us": 229.892, - "pct_cuda_time": 0.8463307670714703, + "cpu_time_us": 93.289, + "cuda_time_us": 248.511, + "pct_cuda_time": 0.89570659435781, "trace": "" }, "children": [ @@ -5921,8 +5921,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 229.892, - "pct_cuda_time": 0.8463307670714703, + "cuda_time_us": 248.511, + "pct_cuda_time": 0.89570659435781, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -5932,9 +5932,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.999, - "cuda_time_us": 23.555, - "pct_cuda_time": 0.08671602847584292, + "cpu_time_us": 26.449, + "cuda_time_us": 25.176, + "pct_cuda_time": 0.09074169441011554, "trace": "" }, "children": [ @@ -5942,8 +5942,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 23.555, - "pct_cuda_time": 0.08671602847584292, + "cuda_time_us": 25.176, + "pct_cuda_time": 0.09074169441011554, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -5953,9 +5953,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 82.653, - "cuda_time_us": 138.886, - "pct_cuda_time": 0.5112987616597716, + "cpu_time_us": 82.774, + "cuda_time_us": 158.146, + "pct_cuda_time": 0.570004607728874, "trace": "" }, "children": [ @@ -5963,8 +5963,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 138.886, - "pct_cuda_time": 0.5112987616597716, + "cuda_time_us": 158.146, + "pct_cuda_time": 0.570004607728874, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -5978,18 +5978,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1042.416, - "cuda_time_us": 771.6580000000001, - "pct_cuda_time": 2.8408031034435157, + "cpu_time_us": 1058.111, + "cuda_time_us": 831.759, + "pct_cuda_time": 2.9979035986996863, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.48, - "cuda_time_us": 15.342, - "pct_cuda_time": 0.05648046312359933, + "cpu_time_us": 33.81, + "cuda_time_us": 15.422, + "pct_cuda_time": 0.05558541512523046, "trace": "" }, "children": [ @@ -5997,8 +5997,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.342, - "pct_cuda_time": 0.05648046312359933, + "cuda_time_us": 15.422, + "pct_cuda_time": 0.05558541512523046, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6008,18 +6008,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 739.634, - "cuda_time_us": 315.58200000000005, - "pct_cuda_time": 1.1617923030551247, + "cpu_time_us": 747.598, + "cuda_time_us": 333.009, + "pct_cuda_time": 1.2002621907299875, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 83.915, - "cuda_time_us": 66.338, - "pct_cuda_time": 0.24421854795289608, + "cpu_time_us": 91.307, + "cuda_time_us": 67.54, + "pct_cuda_time": 0.24343398635443295, "trace": "" }, "children": [ @@ -6027,8 +6027,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 66.338, - "pct_cuda_time": 0.24421854795289608, + "cuda_time_us": 67.54, + "pct_cuda_time": 0.24343398635443295, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -6038,9 +6038,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 80.4, - "cuda_time_us": 39.934, - "pct_cuda_time": 0.14701413208042075, + "cpu_time_us": 75.392, + "cuda_time_us": 39.734, + "pct_cuda_time": 0.14321299990830677, "trace": "" }, "children": [ @@ -6048,8 +6048,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.411, - "pct_cuda_time": 0.030964488028457433, + "cuda_time_us": 7.931, + "pct_cuda_time": 0.02858565214357429, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -6058,8 +6058,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.568, - "pct_cuda_time": 0.020498189197770893, + "cuda_time_us": 5.246, + "pct_cuda_time": 0.01890812396232389, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -6068,8 +6068,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.416, - "pct_cuda_time": 0.04202717813968257, + "cuda_time_us": 11.977, + "pct_cuda_time": 0.04316862384612146, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -6078,8 +6078,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.812, - "pct_cuda_time": 0.032440740519173325, + "cuda_time_us": 9.413, + "pct_cuda_time": 0.03392721518439854, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -6088,8 +6088,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.727, - "pct_cuda_time": 0.021083536195336552, + "cuda_time_us": 5.167, + "pct_cuda_time": 0.018623384771888584, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -6099,9 +6099,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 437.513, - "cuda_time_us": 162.722, - "pct_cuda_time": 0.5990492713074129, + "cpu_time_us": 443.282, + "cuda_time_us": 159.43599999999998, + "pct_cuda_time": 0.574654146408134, "trace": "" }, "children": [ @@ -6109,8 +6109,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 10.095, - "pct_cuda_time": 0.037164012203932686, + "cuda_time_us": 9.854, + "pct_cuda_time": 0.03551670863986648, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -6119,8 +6119,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 152.627, - "pct_cuda_time": 0.5618852591034803, + "cuda_time_us": 149.582, + "pct_cuda_time": 0.5391374377682675, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -6130,9 +6130,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 102.083, - "cuda_time_us": 46.588, - "pct_cuda_time": 0.17151035171439483, + "cpu_time_us": 97.877, + "cuda_time_us": 66.299, + "pct_cuda_time": 0.23896105805911388, "trace": "" }, "children": [ @@ -6140,8 +6140,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 46.588, - "pct_cuda_time": 0.17151035171439483, + "cuda_time_us": 66.299, + "pct_cuda_time": 0.23896105805911388, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -6153,9 +6153,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.897, - "cuda_time_us": 15.062, - "pct_cuda_time": 0.055449663379458544, + "cpu_time_us": 37.687, + "cuda_time_us": 15.181, + "pct_cuda_time": 0.05471678037972529, "trace": "" }, "children": [ @@ -6163,8 +6163,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.062, - "pct_cuda_time": 0.055449663379458544, + "cuda_time_us": 15.181, + "pct_cuda_time": 0.05471678037972529, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6174,9 +6174,9 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 209.993, - "cuda_time_us": 425.672, - "pct_cuda_time": 1.5670806738853325, + "cpu_time_us": 217.084, + "cuda_time_us": 468.14699999999993, + "pct_cuda_time": 1.6873392124647426, "trace": "" }, "children": [ @@ -6184,8 +6184,8 @@ "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 83.315, - "cuda_time_us": 252.966, - "pct_cuda_time": 0.9312760288439856, + "cuda_time_us": 273.989, + "pct_cuda_time": 0.9875367854199693, "trace": "" }, "children": [ @@ -6193,8 +6193,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 252.966, - "pct_cuda_time": 0.9312760288439856, + "cuda_time_us": 273.989, + "pct_cuda_time": 0.9875367854199693, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -6204,9 +6204,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 28.042, - "cuda_time_us": 26.349, - "pct_cuda_time": 0.09700193735130481, + "cpu_time_us": 33.48, + "cuda_time_us": 25.988, + "pct_cuda_time": 0.09366838077256445, "trace": "" }, "children": [ @@ -6214,8 +6214,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 26.349, - "pct_cuda_time": 0.09700193735130481, + "cuda_time_us": 25.988, + "pct_cuda_time": 0.09366838077256445, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -6225,9 +6225,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 81.442, - "cuda_time_us": 146.357, - "pct_cuda_time": 0.5388027076900421, + "cpu_time_us": 81.973, + "cuda_time_us": 168.17, + "pct_cuda_time": 0.6061340462722088, "trace": "" }, "children": [ @@ -6235,8 +6235,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 146.357, - "pct_cuda_time": 0.5388027076900421, + "cuda_time_us": 168.17, + "pct_cuda_time": 0.6061340462722088, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -6250,18 +6250,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1039.221, - "cuda_time_us": 784.8879999999999, - "pct_cuda_time": 2.8895083913541666, + "cpu_time_us": 1092.023, + "cuda_time_us": 838.969, + "pct_cuda_time": 3.0238905551938444, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.2, - "cuda_time_us": 15.462, - "pct_cuda_time": 0.056922234442516796, + "cpu_time_us": 32.469, + "cuda_time_us": 15.863, + "pct_cuda_time": 0.0571749085806984, "trace": "" }, "children": [ @@ -6269,8 +6269,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.462, - "pct_cuda_time": 0.056922234442516796, + "cuda_time_us": 15.863, + "pct_cuda_time": 0.0571749085806984, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6280,18 +6280,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 729.519, - "cuda_time_us": 313.37899999999996, - "pct_cuda_time": 1.1536821179253312, + "cpu_time_us": 772.896, + "cuda_time_us": 335.952, + "pct_cuda_time": 1.21086962664709, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 80.09, - "cuda_time_us": 66.819, - "pct_cuda_time": 0.24598931465622367, + "cpu_time_us": 94.261, + "cuda_time_us": 76.133, + "pct_cuda_time": 0.27440568082798406, "trace": "" }, "children": [ @@ -6299,8 +6299,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 66.819, - "pct_cuda_time": 0.24598931465622367, + "cuda_time_us": 76.133, + "pct_cuda_time": 0.27440568082798406, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -6310,9 +6310,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 74.061, - "cuda_time_us": 41.657000000000004, - "pct_cuda_time": 0.1533572319345442, + "cpu_time_us": 74.211, + "cuda_time_us": 39.494, + "pct_cuda_time": 0.14234796945635142, "trace": "" }, "children": [ @@ -6320,8 +6320,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.171, - "pct_cuda_time": 0.030080945390622477, + "cuda_time_us": 7.931, + "pct_cuda_time": 0.02858565214357429, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -6330,8 +6330,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.327, - "pct_cuda_time": 0.019610965132278297, + "cuda_time_us": 5.166, + "pct_cuda_time": 0.018619780478338774, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -6340,8 +6340,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 12.938, - "pct_cuda_time": 0.04763031103461922, + "cuda_time_us": 12.217, + "pct_cuda_time": 0.044033654298076805, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -6350,8 +6350,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 9.694, - "pct_cuda_time": 0.035687759713216784, + "cuda_time_us": 8.893, + "pct_cuda_time": 0.0320529825384953, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -6360,8 +6360,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.527, - "pct_cuda_time": 0.020347250663807423, + "cuda_time_us": 5.287, + "pct_cuda_time": 0.01905589999786626, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -6371,9 +6371,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 439.677, - "cuda_time_us": 159.636, - "pct_cuda_time": 0.5876883855559184, + "cpu_time_us": 464.865, + "cuda_time_us": 153.386, + "pct_cuda_time": 0.5528481704317597, "trace": "" }, "children": [ @@ -6381,8 +6381,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 10.254, - "pct_cuda_time": 0.037749359201498335, + "cuda_time_us": 9.893, + "pct_cuda_time": 0.03565727608830923, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -6391,8 +6391,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 149.382, - "pct_cuda_time": 0.5499390263544202, + "cuda_time_us": 143.493, + "pct_cuda_time": 0.5171908943434506, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -6402,9 +6402,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 100.229, - "cuda_time_us": 45.267, - "pct_cuda_time": 0.16664718577864493, + "cpu_time_us": 99.489, + "cuda_time_us": 66.939, + "pct_cuda_time": 0.24126780593099473, "trace": "" }, "children": [ @@ -6412,8 +6412,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 45.267, - "pct_cuda_time": 0.16664718577864493, + "cuda_time_us": 66.939, + "pct_cuda_time": 0.24126780593099473, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -6425,9 +6425,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.358, - "cuda_time_us": 15.061, - "pct_cuda_time": 0.0554459819518009, + "cpu_time_us": 38.478, + "cuda_time_us": 14.18, + "pct_cuda_time": 0.05110888253636155, "trace": "" }, "children": [ @@ -6435,8 +6435,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.061, - "pct_cuda_time": 0.0554459819518009, + "cuda_time_us": 14.18, + "pct_cuda_time": 0.05110888253636155, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6446,18 +6446,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 212.728, - "cuda_time_us": 440.986, - "pct_cuda_time": 1.6234580570345176, + "cpu_time_us": 226.629, + "cuda_time_us": 472.974, + "pct_cuda_time": 1.7047371374296947, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 86.93, - "cuda_time_us": 254.88, - "pct_cuda_time": 0.9383222813807194, + "cpu_time_us": 93.4, + "cuda_time_us": 274.169, + "pct_cuda_time": 0.9881855582589357, "trace": "" }, "children": [ @@ -6465,8 +6465,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 254.88, - "pct_cuda_time": 0.9383222813807194, + "cuda_time_us": 274.169, + "pct_cuda_time": 0.9881855582589357, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -6476,9 +6476,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.8, - "cuda_time_us": 26.298, - "pct_cuda_time": 0.09681418454076489, + "cpu_time_us": 25.889, + "cuda_time_us": 25.757, + "pct_cuda_time": 0.09283578896255745, "trace": "" }, "children": [ @@ -6486,8 +6486,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 26.298, - "pct_cuda_time": 0.09681418454076489, + "cuda_time_us": 25.757, + "pct_cuda_time": 0.09283578896255745, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -6497,9 +6497,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 81.602, - "cuda_time_us": 159.808, - "pct_cuda_time": 0.5883215911130335, + "cpu_time_us": 89.534, + "cuda_time_us": 173.048, + "pct_cuda_time": 0.6237157902082012, "trace": "" }, "children": [ @@ -6507,8 +6507,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 159.808, - "pct_cuda_time": 0.5883215911130335, + "cuda_time_us": 173.048, + "pct_cuda_time": 0.6237157902082012, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -6522,18 +6522,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1026.603, - "cuda_time_us": 798.758, - "pct_cuda_time": 2.940569792965712, + "cpu_time_us": 1050.971, + "cuda_time_us": 847.874, + "pct_cuda_time": 3.0559867892549377, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.57, - "cuda_time_us": 16.022, - "pct_cuda_time": 0.05898383393079835, + "cpu_time_us": 32.729, + "cuda_time_us": 15.943, + "pct_cuda_time": 0.057463252064683515, "trace": "" }, "children": [ @@ -6541,8 +6541,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.022, - "pct_cuda_time": 0.05898383393079835, + "cuda_time_us": 15.943, + "pct_cuda_time": 0.057463252064683515, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6552,18 +6552,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 722.018, - "cuda_time_us": 314.331, - "pct_cuda_time": 1.1571868370554101, + "cpu_time_us": 730.211, + "cuda_time_us": 327.982, + "pct_cuda_time": 1.1821434070550731, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 82.693, - "cuda_time_us": 70.455, - "pct_cuda_time": 0.2593749856194232, + "cpu_time_us": 80.591, + "cuda_time_us": 76.573, + "pct_cuda_time": 0.27599156998990215, "trace": "" }, "children": [ @@ -6571,8 +6571,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 70.455, - "pct_cuda_time": 0.2593749856194232, + "cuda_time_us": 76.573, + "pct_cuda_time": 0.27599156998990215, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -6582,9 +6582,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 73.95, - "cuda_time_us": 40.013999999999996, - "pct_cuda_time": 0.1473086462930324, + "cpu_time_us": 74.331, + "cuda_time_us": 38.253, + "pct_cuda_time": 0.13787504116103233, "trace": "" }, "children": [ @@ -6592,8 +6592,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.89, - "pct_cuda_time": 0.029046464218824057, + "cuda_time_us": 7.691, + "pct_cuda_time": 0.02772062169161895, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -6602,8 +6602,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.488, - "pct_cuda_time": 0.020203674985159244, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.019488415223843928, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -6612,8 +6612,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 12.658, - "pct_cuda_time": 0.04659951129047844, + "cuda_time_us": 10.975, + "pct_cuda_time": 0.0395571217092079, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -6622,8 +6622,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.571, - "pct_cuda_time": 0.031553516453680736, + "cuda_time_us": 9.093, + "pct_cuda_time": 0.032773841248458085, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -6632,8 +6632,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.407, - "pct_cuda_time": 0.019905479344889947, + "cuda_time_us": 5.087, + "pct_cuda_time": 0.01833504128790347, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -6643,9 +6643,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 430.372, - "cuda_time_us": 157.433, - "pct_cuda_time": 0.5795782004261252, + "cpu_time_us": 439.287, + "cuda_time_us": 147.899, + "pct_cuda_time": 0.5330714117239307, "trace": "" }, "children": [ @@ -6653,8 +6653,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 10.054, - "pct_cuda_time": 0.03701307366996921, + "cuda_time_us": 9.773, + "pct_cuda_time": 0.035224760862331554, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -6663,8 +6663,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 147.379, - "pct_cuda_time": 0.5425651267561559, + "cuda_time_us": 138.126, + "pct_cuda_time": 0.4978466508615991, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -6674,9 +6674,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 100.249, - "cuda_time_us": 46.429, - "pct_cuda_time": 0.17092500471682917, + "cpu_time_us": 99.799, + "cuda_time_us": 65.257, + "pct_cuda_time": 0.23520538418020775, "trace": "" }, "children": [ @@ -6684,8 +6684,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 46.429, - "pct_cuda_time": 0.17092500471682917, + "cuda_time_us": 65.257, + "pct_cuda_time": 0.23520538418020775, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -6697,9 +6697,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.319, - "cuda_time_us": 14.66, - "pct_cuda_time": 0.053969729461085005, + "cpu_time_us": 38.628, + "cuda_time_us": 15.382, + "pct_cuda_time": 0.0554412433832379, "trace": "" }, "children": [ @@ -6707,8 +6707,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.66, - "pct_cuda_time": 0.053969729461085005, + "cuda_time_us": 15.382, + "pct_cuda_time": 0.0554412433832379, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6718,18 +6718,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 210.634, - "cuda_time_us": 453.745, - "pct_cuda_time": 1.6704293925184184, + "cpu_time_us": 224.816, + "cuda_time_us": 488.567, + "pct_cuda_time": 1.760938886751943, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 85.157, - "cuda_time_us": 261.46, - "pct_cuda_time": 0.9625460753680275, + "cpu_time_us": 84.055, + "cuda_time_us": 284.364, + "pct_cuda_time": 1.024931330999289, "trace": "" }, "children": [ @@ -6737,8 +6737,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 261.46, - "pct_cuda_time": 0.9625460753680275, + "cuda_time_us": 284.364, + "pct_cuda_time": 1.024931330999289, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -6748,9 +6748,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 27.34, - "cuda_time_us": 25.867, - "pct_cuda_time": 0.09522748922031964, + "cpu_time_us": 26.71, + "cuda_time_us": 25.166, + "pct_cuda_time": 0.0907056514746174, "trace": "" }, "children": [ @@ -6758,8 +6758,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 25.867, - "pct_cuda_time": 0.09522748922031964, + "cuda_time_us": 25.166, + "pct_cuda_time": 0.0907056514746174, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -6769,9 +6769,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 81.142, - "cuda_time_us": 166.418, - "pct_cuda_time": 0.6126558279300712, + "cpu_time_us": 94.611, + "cuda_time_us": 179.037, + "pct_cuda_time": 0.6453019042780369, "trace": "" }, "children": [ @@ -6779,8 +6779,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 166.418, - "pct_cuda_time": 0.6126558279300712, + "cuda_time_us": 179.037, + "pct_cuda_time": 0.6453019042780369, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -6794,18 +6794,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1048.806, - "cuda_time_us": 820.4200000000001, - "pct_cuda_time": 3.0203168788856316, + "cpu_time_us": 1152.373, + "cuda_time_us": 852.84, + "pct_cuda_time": 3.0738857110233138, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.199, - "cuda_time_us": 14.22, - "pct_cuda_time": 0.05234990129172093, + "cpu_time_us": 40.58, + "cuda_time_us": 15.182, + "pct_cuda_time": 0.054720384673275106, "trace": "" }, "children": [ @@ -6813,8 +6813,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.22, - "pct_cuda_time": 0.05234990129172093, + "cuda_time_us": 15.182, + "pct_cuda_time": 0.054720384673275106, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6824,18 +6824,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 734.637, - "cuda_time_us": 317.545, - "pct_cuda_time": 1.1690189455470832, + "cpu_time_us": 828.058, + "cuda_time_us": 334.771, + "pct_cuda_time": 1.2066129559647598, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 82.473, - "cuda_time_us": 75.832, - "pct_cuda_time": 0.27917002213458375, + "cpu_time_us": 96.665, + "cuda_time_us": 81.04, + "pct_cuda_time": 0.29209194927692106, "trace": "" }, "children": [ @@ -6843,8 +6843,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 75.832, - "pct_cuda_time": 0.27917002213458375, + "cuda_time_us": 81.04, + "pct_cuda_time": 0.29209194927692106, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -6854,9 +6854,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 74.492, - "cuda_time_us": 38.653, - "pct_cuda_time": 0.1422982232509767, + "cpu_time_us": 77.106, + "cuda_time_us": 37.891999999999996, + "pct_cuda_time": 0.1365738911895495, "trace": "" }, "children": [ @@ -6864,8 +6864,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.011, - "pct_cuda_time": 0.029491916965399174, + "cuda_time_us": 7.731, + "pct_cuda_time": 0.027864793433611507, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -6874,8 +6874,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.247, - "pct_cuda_time": 0.019316450919666644, + "cuda_time_us": 4.966, + "pct_cuda_time": 0.017898921768375986, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -6884,8 +6884,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.176, - "pct_cuda_time": 0.04114363550184761, + "cuda_time_us": 11.456, + "pct_cuda_time": 0.0412907869066684, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -6894,8 +6894,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.772, - "pct_cuda_time": 0.03229348341286751, + "cuda_time_us": 8.692, + "pct_cuda_time": 0.03132851953498269, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -6904,8 +6904,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.447, - "pct_cuda_time": 0.02005273645119577, + "cuda_time_us": 5.047, + "pct_cuda_time": 0.018190869545910912, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -6915,9 +6915,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 441.108, - "cuda_time_us": 155.63, - "pct_cuda_time": 0.5729405863593902, + "cpu_time_us": 516.783, + "cuda_time_us": 151.704, + "pct_cuda_time": 0.5467857486809727, "trace": "" }, "children": [ @@ -6925,8 +6925,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 10.214, - "pct_cuda_time": 0.03760210209519251, + "cuda_time_us": 10.054, + "pct_cuda_time": 0.03623756734982927, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -6935,8 +6935,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 145.416, - "pct_cuda_time": 0.5353384842641975, + "cuda_time_us": 141.65, + "pct_cuda_time": 0.5105481813311434, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -6946,9 +6946,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 101.361, - "cuda_time_us": 47.43, - "pct_cuda_time": 0.17461011380213246, + "cpu_time_us": 100.441, + "cuda_time_us": 64.135, + "pct_cuda_time": 0.23116136681731653, "trace": "" }, "children": [ @@ -6956,8 +6956,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 47.43, - "pct_cuda_time": 0.17461011380213246, + "cuda_time_us": 64.135, + "pct_cuda_time": 0.23116136681731653, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -6969,9 +6969,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.076, - "cuda_time_us": 14.701, - "pct_cuda_time": 0.05412066799504847, + "cpu_time_us": 37.877, + "cuda_time_us": 14.58, + "pct_cuda_time": 0.05255059995628713, "trace": "" }, "children": [ @@ -6979,8 +6979,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.701, - "pct_cuda_time": 0.05412066799504847, + "cuda_time_us": 14.58, + "pct_cuda_time": 0.05255059995628713, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6990,18 +6990,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 222.102, - "cuda_time_us": 473.95399999999995, - "pct_cuda_time": 1.7448273640517789, + "cpu_time_us": 222.472, + "cuda_time_us": 488.307, + "pct_cuda_time": 1.7600017704289916, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 84.696, - "cuda_time_us": 277.212, - "pct_cuda_time": 1.0205359238312615, + "cpu_time_us": 82.614, + "cuda_time_us": 284.374, + "pct_cuda_time": 1.024967373934787, "trace": "" }, "children": [ @@ -7009,8 +7009,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 277.212, - "pct_cuda_time": 1.0205359238312615, + "cuda_time_us": 284.374, + "pct_cuda_time": 1.024967373934787, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -7020,9 +7020,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 27.361, - "cuda_time_us": 27.06, - "pct_cuda_time": 0.09961943241589086, + "cpu_time_us": 26.239, + "cuda_time_us": 25.718, + "pct_cuda_time": 0.09269522151411469, "trace": "" }, "children": [ @@ -7030,8 +7030,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 27.06, - "pct_cuda_time": 0.09961943241589086, + "cuda_time_us": 25.718, + "pct_cuda_time": 0.09269522151411469, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -7041,9 +7041,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 91.977, - "cuda_time_us": 169.682, - "pct_cuda_time": 0.6246720078046265, + "cpu_time_us": 95.272, + "cuda_time_us": 178.215, + "pct_cuda_time": 0.6423391749800899, "trace": "" }, "children": [ @@ -7051,8 +7051,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 169.682, - "pct_cuda_time": 0.6246720078046265, + "cuda_time_us": 178.215, + "pct_cuda_time": 0.6423391749800899, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -7066,18 +7066,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1027.384, - "cuda_time_us": 819.6600000000001, - "pct_cuda_time": 3.017518993865821, + "cpu_time_us": 1083.991, + "cuda_time_us": 851.419, + "pct_cuda_time": 3.068764009889028, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.339, - "cuda_time_us": 15.342, - "pct_cuda_time": 0.05648046312359933, + "cpu_time_us": 34.602, + "cuda_time_us": 14.781, + "pct_cuda_time": 0.053275062959799724, "trace": "" }, "children": [ @@ -7085,8 +7085,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.342, - "pct_cuda_time": 0.05648046312359933, + "cuda_time_us": 14.781, + "pct_cuda_time": 0.053275062959799724, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7096,18 +7096,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 725.313, - "cuda_time_us": 311.14500000000004, - "pct_cuda_time": 1.145457808538151, + "cpu_time_us": 777.672, + "cuda_time_us": 326.799, + "pct_cuda_time": 1.177879527785643, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 81.832, - "cuda_time_us": 76.523, - "pct_cuda_time": 0.2817138886460169, + "cpu_time_us": 86.44, + "cuda_time_us": 79.679, + "pct_cuda_time": 0.28718650575562427, "trace": "" }, "children": [ @@ -7115,8 +7115,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 76.523, - "pct_cuda_time": 0.2817138886460169, + "cuda_time_us": 79.679, + "pct_cuda_time": 0.28718650575562427, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -7126,9 +7126,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 76.033, - "cuda_time_us": 38.211999999999996, - "pct_cuda_time": 0.14067471365395498, + "cpu_time_us": 84.446, + "cuda_time_us": 38.132000000000005, + "pct_cuda_time": 0.13743892164150487, "trace": "" }, "children": [ @@ -7136,8 +7136,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.771, - "pct_cuda_time": 0.028608374327564228, + "cuda_time_us": 7.53, + "pct_cuda_time": 0.027140330430098905, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -7146,8 +7146,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.367, - "pct_cuda_time": 0.01975822223858412, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.01761057828439087, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -7156,8 +7156,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 10.615, - "pct_cuda_time": 0.039078354585908415, + "cuda_time_us": 12.338, + "pct_cuda_time": 0.04446977381760429, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -7166,8 +7166,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 9.132, - "pct_cuda_time": 0.03361879736961994, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.029883197821507307, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -7176,8 +7176,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.327, - "pct_cuda_time": 0.019610965132278297, + "cuda_time_us": 5.087, + "pct_cuda_time": 0.01833504128790347, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -7187,9 +7187,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 431.954, - "cuda_time_us": 150.663, - "pct_cuda_time": 0.5546549351838643, + "cpu_time_us": 443.092, + "cuda_time_us": 145.094, + "pct_cuda_time": 0.5229613683167026, "trace": "" }, "children": [ @@ -7197,8 +7197,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 11.256, - "pct_cuda_time": 0.04143814971445926, + "cuda_time_us": 9.533, + "pct_cuda_time": 0.03435973041037621, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -7207,8 +7207,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 139.407, - "pct_cuda_time": 0.513216785469405, + "cuda_time_us": 135.561, + "pct_cuda_time": 0.48860163790632644, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -7218,9 +7218,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.949, - "cuda_time_us": 45.747, - "pct_cuda_time": 0.16841427105431483, + "cpu_time_us": 128.031, + "cuda_time_us": 63.894, + "pct_cuda_time": 0.23029273207181133, "trace": "" }, "children": [ @@ -7228,8 +7228,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 45.747, - "pct_cuda_time": 0.16841427105431483, + "cuda_time_us": 63.894, + "pct_cuda_time": 0.23029273207181133, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -7241,9 +7241,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.736, - "cuda_time_us": 15.422, - "pct_cuda_time": 0.05677497733621098, + "cpu_time_us": 38.518, + "cuda_time_us": 14.742, + "pct_cuda_time": 0.05313449551135698, "trace": "" }, "children": [ @@ -7251,8 +7251,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.422, - "pct_cuda_time": 0.05677497733621098, + "cuda_time_us": 14.742, + "pct_cuda_time": 0.05313449551135698, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7262,18 +7262,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 211.696, - "cuda_time_us": 477.75100000000003, - "pct_cuda_time": 1.7588057448678596, + "cpu_time_us": 211.476, + "cuda_time_us": 495.097, + "pct_cuda_time": 1.7844749236322281, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.694, - "cuda_time_us": 277.023, - "pct_cuda_time": 1.0198401340039667, + "cpu_time_us": 83.906, + "cuda_time_us": 289.291, + "pct_cuda_time": 1.0426896853192222, "trace": "" }, "children": [ @@ -7281,8 +7281,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 277.023, - "pct_cuda_time": 1.0198401340039667, + "cuda_time_us": 289.291, + "pct_cuda_time": 1.0426896853192222, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -7292,9 +7292,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.81, - "cuda_time_us": 25.517, - "pct_cuda_time": 0.09393898954014365, + "cpu_time_us": 25.999, + "cuda_time_us": 24.876, + "pct_cuda_time": 0.08966040634517136, "trace": "" }, "children": [ @@ -7302,8 +7302,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 25.517, - "pct_cuda_time": 0.09393898954014365, + "cuda_time_us": 24.876, + "pct_cuda_time": 0.08966040634517136, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -7313,9 +7313,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 83.995, - "cuda_time_us": 175.211, - "pct_cuda_time": 0.6450266213237493, + "cpu_time_us": 83.515, + "cuda_time_us": 180.93, + "pct_cuda_time": 0.6521248319678347, "trace": "" }, "children": [ @@ -7323,8 +7323,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 175.211, - "pct_cuda_time": 0.6450266213237493, + "cuda_time_us": 180.93, + "pct_cuda_time": 0.6521248319678347, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -7338,18 +7338,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1047.404, - "cuda_time_us": 825.867, - "pct_cuda_time": 3.040369615336827, + "cpu_time_us": 1071.512, + "cuda_time_us": 852.5700000000002, + "pct_cuda_time": 3.072912551764864, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.3, - "cuda_time_us": 15.662, - "pct_cuda_time": 0.05765851997404593, + "cpu_time_us": 33.981, + "cuda_time_us": 15.261, + "pct_cuda_time": 0.055005123863710405, "trace": "" }, "children": [ @@ -7357,8 +7357,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.662, - "pct_cuda_time": 0.05765851997404593, + "cuda_time_us": 15.261, + "pct_cuda_time": 0.055005123863710405, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7368,18 +7368,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 738.763, - "cuda_time_us": 312.937, - "pct_cuda_time": 1.152054926900652, + "cpu_time_us": 756.921, + "cuda_time_us": 331.53700000000003, + "pct_cuda_time": 1.1949566706246617, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 91.016, - "cuda_time_us": 79.417, - "pct_cuda_time": 0.29236794028724333, + "cpu_time_us": 80.861, + "cuda_time_us": 80.209, + "pct_cuda_time": 0.28909678133702565, "trace": "" }, "children": [ @@ -7387,8 +7387,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 79.417, - "pct_cuda_time": 0.29236794028724333, + "cuda_time_us": 80.209, + "pct_cuda_time": 0.28909678133702565, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -7398,9 +7398,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 81.412, - "cuda_time_us": 38.132, - "pct_cuda_time": 0.14038019944134333, + "cpu_time_us": 82.003, + "cuda_time_us": 38.132000000000005, + "pct_cuda_time": 0.13743892164150487, "trace": "" }, "children": [ @@ -7408,8 +7408,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.931, - "pct_cuda_time": 0.029197402752787528, + "cuda_time_us": 7.731, + "pct_cuda_time": 0.027864793433611507, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -7418,8 +7418,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.087, - "pct_cuda_time": 0.01872742249444334, + "cuda_time_us": 5.287, + "pct_cuda_time": 0.01905589999786626, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -7428,8 +7428,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.296, - "pct_cuda_time": 0.041585406820765086, + "cuda_time_us": 11.576, + "pct_cuda_time": 0.04172330213264608, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -7438,8 +7438,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.572, - "pct_cuda_time": 0.03155719788133837, + "cuda_time_us": 8.532, + "pct_cuda_time": 0.030751832567012465, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -7448,8 +7448,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.246, - "pct_cuda_time": 0.019312769492009, + "cuda_time_us": 5.006, + "pct_cuda_time": 0.018043093510368542, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -7459,9 +7459,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 429.02, - "cuda_time_us": 149.621, - "pct_cuda_time": 0.5508188875645975, + "cpu_time_us": 458.325, + "cuda_time_us": 148.5, + "pct_cuda_time": 0.5352375921473689, "trace": "" }, "children": [ @@ -7469,8 +7469,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.734, - "pct_cuda_time": 0.03583501681952261, + "cuda_time_us": 9.814, + "pct_cuda_time": 0.035372536897873924, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -7479,8 +7479,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 139.887, - "pct_cuda_time": 0.5149838707450749, + "cuda_time_us": 138.686, + "pct_cuda_time": 0.4998650552494949, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -7490,9 +7490,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 100.78, - "cuda_time_us": 45.767, - "pct_cuda_time": 0.16848789960746777, + "cpu_time_us": 101.322, + "cuda_time_us": 64.696, + "pct_cuda_time": 0.2331833754987621, "trace": "" }, "children": [ @@ -7500,8 +7500,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 45.767, - "pct_cuda_time": 0.16848789960746777, + "cuda_time_us": 64.696, + "pct_cuda_time": 0.2331833754987621, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -7513,9 +7513,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.337, - "cuda_time_us": 15.101, - "pct_cuda_time": 0.05559323905810672, + "cpu_time_us": 46.429, + "cuda_time_us": 15.663, + "pct_cuda_time": 0.05645404987073561, "trace": "" }, "children": [ @@ -7523,8 +7523,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.101, - "pct_cuda_time": 0.05559323905810672, + "cuda_time_us": 15.663, + "pct_cuda_time": 0.05645404987073561, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7534,18 +7534,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 216.383, - "cuda_time_us": 482.16700000000003, - "pct_cuda_time": 1.7750629294040228, + "cpu_time_us": 213.129, + "cuda_time_us": 490.10900000000004, + "pct_cuda_time": 1.7664967074057563, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 85.187, - "cuda_time_us": 281.72, - "pct_cuda_time": 1.0371317997119283, + "cpu_time_us": 84.587, + "cuda_time_us": 284.264, + "pct_cuda_time": 1.0245709016443074, "trace": "" }, "children": [ @@ -7553,8 +7553,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 281.72, - "pct_cuda_time": 1.0371317997119283, + "cuda_time_us": 284.264, + "pct_cuda_time": 1.0245709016443074, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -7564,9 +7564,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.518, - "cuda_time_us": 25.868, - "pct_cuda_time": 0.09523117064797726, + "cpu_time_us": 27.04, + "cuda_time_us": 25.396, + "pct_cuda_time": 0.09153463899107461, "trace": "" }, "children": [ @@ -7574,8 +7574,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 25.868, - "pct_cuda_time": 0.09523117064797726, + "cuda_time_us": 25.396, + "pct_cuda_time": 0.09153463899107461, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -7585,9 +7585,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 86.469, - "cuda_time_us": 174.579, - "pct_cuda_time": 0.6426999590441173, + "cpu_time_us": 83.165, + "cuda_time_us": 180.449, + "pct_cuda_time": 0.6503911667703742, "trace": "" }, "children": [ @@ -7595,8 +7595,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 174.579, - "pct_cuda_time": 0.6426999590441173, + "cuda_time_us": 180.449, + "pct_cuda_time": 0.6503911667703742, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -7610,18 +7610,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1028.906, - "cuda_time_us": 822.806, - "pct_cuda_time": 3.029100765276774, + "cpu_time_us": 1093.745, + "cuda_time_us": 850.1379999999999, + "pct_cuda_time": 3.0641469098517162, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.681, - "cuda_time_us": 14.06, - "pct_cuda_time": 0.05176087286649762, + "cpu_time_us": 40.671, + "cuda_time_us": 15.062, + "pct_cuda_time": 0.05428786944729744, "trace": "" }, "children": [ @@ -7629,8 +7629,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.06, - "pct_cuda_time": 0.05176087286649762, + "cuda_time_us": 15.062, + "pct_cuda_time": 0.05428786944729744, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7640,18 +7640,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 727.546, - "cuda_time_us": 313.4, - "pct_cuda_time": 1.1537594279061418, + "cpu_time_us": 757.622, + "cuda_time_us": 332.41799999999995, + "pct_cuda_time": 1.1981320532420474, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 85.718, - "cuda_time_us": 81.851, - "pct_cuda_time": 0.3013285352059528, + "cpu_time_us": 89.454, + "cuda_time_us": 80.449, + "pct_cuda_time": 0.289961811788981, "trace": "" }, "children": [ @@ -7659,8 +7659,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 81.851, - "pct_cuda_time": 0.3013285352059528, + "cuda_time_us": 80.449, + "pct_cuda_time": 0.289961811788981, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -7670,9 +7670,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 75.553, - "cuda_time_us": 37.331, - "pct_cuda_time": 0.1374313758875692, + "cpu_time_us": 76.875, + "cuda_time_us": 38.211999999999996, + "pct_cuda_time": 0.13772726512548994, "trace": "" }, "children": [ @@ -7680,8 +7680,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.011, - "pct_cuda_time": 0.029491916965399174, + "cuda_time_us": 7.891, + "pct_cuda_time": 0.028441480401581735, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -7690,8 +7690,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.967, - "pct_cuda_time": 0.018285651175525865, + "cuda_time_us": 4.926, + "pct_cuda_time": 0.017754750026383426, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -7700,8 +7700,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 10.214, - "pct_cuda_time": 0.03760210209519251, + "cuda_time_us": 11.296, + "pct_cuda_time": 0.04071409993869817, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -7711,7 +7711,7 @@ "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 8.892, - "pct_cuda_time": 0.03273525473178498, + "pct_cuda_time": 0.03204937824494548, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -7720,8 +7720,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.247, - "pct_cuda_time": 0.019316450919666644, + "cuda_time_us": 5.207, + "pct_cuda_time": 0.018767556513881144, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -7731,9 +7731,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 430.533, - "cuda_time_us": 149.302, - "pct_cuda_time": 0.5496445121418084, + "cpu_time_us": 451.645, + "cuda_time_us": 149.903, + "pct_cuda_time": 0.5402944159977577, "trace": "" }, "children": [ @@ -7742,7 +7742,7 @@ "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 9.614, - "pct_cuda_time": 0.03539324550060513, + "pct_cuda_time": 0.03465167818791114, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -7751,8 +7751,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 139.688, - "pct_cuda_time": 0.5142512666412034, + "cuda_time_us": 140.289, + "pct_cuda_time": 0.5056427378098466, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -7762,9 +7762,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.699, - "cuda_time_us": 44.916, - "pct_cuda_time": 0.1653550046708113, + "cpu_time_us": 102.843, + "cuda_time_us": 63.854, + "pct_cuda_time": 0.2301485603298188, "trace": "" }, "children": [ @@ -7772,8 +7772,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 44.916, - "pct_cuda_time": 0.1653550046708113, + "cuda_time_us": 63.854, + "pct_cuda_time": 0.2301485603298188, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -7785,9 +7785,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 36.675, - "cuda_time_us": 15.463, - "pct_cuda_time": 0.05692591587017445, + "cpu_time_us": 39.639, + "cuda_time_us": 14.501, + "pct_cuda_time": 0.05226586076585182, "trace": "" }, "children": [ @@ -7795,8 +7795,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.463, - "pct_cuda_time": 0.05692591587017445, + "cuda_time_us": 14.501, + "pct_cuda_time": 0.05226586076585182, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7806,18 +7806,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 210.404, - "cuda_time_us": 479.88300000000004, - "pct_cuda_time": 1.7666545486339602, + "cpu_time_us": 231.095, + "cuda_time_us": 488.15700000000004, + "pct_cuda_time": 1.7594611263965196, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.605, - "cuda_time_us": 277.723, - "pct_cuda_time": 1.0224171333643186, + "cpu_time_us": 96.264, + "cuda_time_us": 284.935, + "pct_cuda_time": 1.0269893826162326, "trace": "" }, "children": [ @@ -7825,8 +7825,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 277.723, - "pct_cuda_time": 1.0224171333643186, + "cuda_time_us": 284.935, + "pct_cuda_time": 1.0269893826162326, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -7836,9 +7836,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.64, - "cuda_time_us": 26.378, - "pct_cuda_time": 0.09710869875337655, + "cpu_time_us": 27.691, + "cuda_time_us": 25.557, + "pct_cuda_time": 0.09211493025259465, "trace": "" }, "children": [ @@ -7846,8 +7846,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 26.378, - "pct_cuda_time": 0.09710869875337655, + "cuda_time_us": 25.557, + "pct_cuda_time": 0.09211493025259465, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -7857,9 +7857,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 82.653, - "cuda_time_us": 175.782, - "pct_cuda_time": 0.6471287165162649, + "cpu_time_us": 87.2, + "cuda_time_us": 177.665, + "pct_cuda_time": 0.6403568135276921, "trace": "" }, "children": [ @@ -7867,8 +7867,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 175.782, - "pct_cuda_time": 0.6471287165162649, + "cuda_time_us": 177.665, + "pct_cuda_time": 0.6403568135276921, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -7882,18 +7882,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1035.786, - "cuda_time_us": 831.086, - "pct_cuda_time": 3.05958298628208, + "cpu_time_us": 1102.638, + "cuda_time_us": 847.0020000000001, + "pct_cuda_time": 3.0528438452795, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.079, - "cuda_time_us": 15.021, - "pct_cuda_time": 0.05529872484549508, + "cpu_time_us": 34.712, + "cuda_time_us": 14.821, + "pct_cuda_time": 0.05341923470179228, "trace": "" }, "children": [ @@ -7901,8 +7901,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.021, - "pct_cuda_time": 0.05529872484549508, + "cuda_time_us": 14.821, + "pct_cuda_time": 0.05341923470179228, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7912,18 +7912,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 729.268, - "cuda_time_us": 311.20599999999996, - "pct_cuda_time": 1.1456823756252672, + "cpu_time_us": 763.802, + "cuda_time_us": 329.694, + "pct_cuda_time": 1.1883139576123545, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.125, - "cuda_time_us": 79.989, - "pct_cuda_time": 0.2944737169074167, + "cpu_time_us": 85.398, + "cuda_time_us": 82.853, + "pct_cuda_time": 0.29862653348273366, "trace": "" }, "children": [ @@ -7931,8 +7931,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 79.989, - "pct_cuda_time": 0.2944737169074167, + "cuda_time_us": 82.853, + "pct_cuda_time": 0.29862653348273366, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -7942,9 +7942,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 74.061, - "cuda_time_us": 38.492, - "pct_cuda_time": 0.14170551339809576, + "cpu_time_us": 88.903, + "cuda_time_us": 37.21, + "pct_cuda_time": 0.1341157629885764, "trace": "" }, "children": [ @@ -7952,8 +7952,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.65, - "pct_cuda_time": 0.028162921580989105, + "cuda_time_us": 7.81, + "pct_cuda_time": 0.028149532624046805, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -7962,8 +7962,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.287, - "pct_cuda_time": 0.01946370802597247, + "cuda_time_us": 5.087, + "pct_cuda_time": 0.01833504128790347, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -7972,8 +7972,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.616, - "pct_cuda_time": 0.042763463671211685, + "cuda_time_us": 10.695, + "pct_cuda_time": 0.03854791951526, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -7982,8 +7982,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.652, - "pct_cuda_time": 0.031851712093950026, + "cuda_time_us": 8.412, + "pct_cuda_time": 0.0303193173410348, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -7992,8 +7992,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.287, - "pct_cuda_time": 0.01946370802597247, + "cuda_time_us": 5.206, + "pct_cuda_time": 0.01876395222033133, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -8003,9 +8003,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 435.891, - "cuda_time_us": 147.25799999999998, - "pct_cuda_time": 0.5421196740095808, + "cpu_time_us": 449.923, + "cuda_time_us": 148.42000000000002, + "pct_cuda_time": 0.5349492486633838, "trace": "" }, "children": [ @@ -8013,8 +8013,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.854, - "pct_cuda_time": 0.03627678813844008, + "cuda_time_us": 9.413, + "pct_cuda_time": 0.03392721518439854, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -8023,8 +8023,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 137.404, - "pct_cuda_time": 0.5058428858711408, + "cuda_time_us": 139.007, + "pct_cuda_time": 0.5010220334789852, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -8034,9 +8034,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.889, - "cuda_time_us": 45.467, - "pct_cuda_time": 0.16738347131017406, + "cpu_time_us": 102.644, + "cuda_time_us": 61.211, + "pct_cuda_time": 0.2206224124776606, "trace": "" }, "children": [ @@ -8044,8 +8044,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 45.467, - "pct_cuda_time": 0.16738347131017406, + "cuda_time_us": 61.211, + "pct_cuda_time": 0.2206224124776606, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -8057,9 +8057,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.159, - "cuda_time_us": 14.66, - "pct_cuda_time": 0.053969729461085005, + "cpu_time_us": 39.289, + "cuda_time_us": 15.662, + "pct_cuda_time": 0.0564504455771858, "trace": "" }, "children": [ @@ -8067,8 +8067,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.66, - "pct_cuda_time": 0.053969729461085005, + "cuda_time_us": 15.662, + "pct_cuda_time": 0.0564504455771858, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8078,18 +8078,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 213.078, - "cuda_time_us": 490.199, - "pct_cuda_time": 1.8046321563502326, + "cpu_time_us": 237.135, + "cuda_time_us": 486.82500000000005, + "pct_cuda_time": 1.7546602073881674, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 84.296, - "cuda_time_us": 286.117, - "pct_cuda_time": 1.053319037122596, + "cpu_time_us": 90.515, + "cuda_time_us": 282.952, + "pct_cuda_time": 1.0198420685069516, "trace": "" }, "children": [ @@ -8097,8 +8097,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 286.117, - "pct_cuda_time": 1.053319037122596, + "cuda_time_us": 282.952, + "pct_cuda_time": 1.0198420685069516, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -8108,9 +8108,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.48, - "cuda_time_us": 26.308, - "pct_cuda_time": 0.09685099881734135, + "cpu_time_us": 28.322, + "cuda_time_us": 25.547, + "pct_cuda_time": 0.09207888731709651, "trace": "" }, "children": [ @@ -8118,8 +8118,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 26.308, - "pct_cuda_time": 0.09685099881734135, + "cuda_time_us": 25.547, + "pct_cuda_time": 0.09207888731709651, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -8129,9 +8129,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 84.336, - "cuda_time_us": 177.774, - "pct_cuda_time": 0.6544621204102951, + "cpu_time_us": 99.789, + "cuda_time_us": 178.326, + "pct_cuda_time": 0.6427392515641192, "trace": "" }, "children": [ @@ -8139,8 +8139,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 177.774, - "pct_cuda_time": 0.6544621204102951, + "cuda_time_us": 178.326, + "pct_cuda_time": 0.6427392515641192, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -8154,18 +8154,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1031.77, - "cuda_time_us": 818.609, - "pct_cuda_time": 3.013649813397635, + "cpu_time_us": 1083.5, + "cuda_time_us": 846.983, + "pct_cuda_time": 3.052775363702053, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.059, - "cuda_time_us": 14.942, - "pct_cuda_time": 0.05500789206054107, + "cpu_time_us": 34.231, + "cuda_time_us": 14.741, + "pct_cuda_time": 0.05313089121780717, "trace": "" }, "children": [ @@ -8173,8 +8173,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.942, - "pct_cuda_time": 0.05500789206054107, + "cuda_time_us": 14.741, + "pct_cuda_time": 0.05313089121780717, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8184,18 +8184,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 719.234, - "cuda_time_us": 309.96400000000006, - "pct_cuda_time": 1.1411100424744718, + "cpu_time_us": 765.925, + "cuda_time_us": 335.01300000000003, + "pct_cuda_time": 1.2074851950038148, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 80.26, - "cuda_time_us": 77.815, - "pct_cuda_time": 0.2864702931796951, + "cpu_time_us": 83.965, + "cuda_time_us": 82.963, + "pct_cuda_time": 0.2990230057732132, "trace": "" }, "children": [ @@ -8203,8 +8203,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 77.815, - "pct_cuda_time": 0.2864702931796951, + "cuda_time_us": 82.963, + "pct_cuda_time": 0.2990230057732132, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -8214,9 +8214,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 74.752, - "cuda_time_us": 39.414, - "pct_cuda_time": 0.14509978969844503, + "cpu_time_us": 81.913, + "cuda_time_us": 37.972, + "pct_cuda_time": 0.13686223467353462, "trace": "" }, "children": [ @@ -8224,8 +8224,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.61, - "pct_cuda_time": 0.028015664474683278, + "cuda_time_us": 7.971, + "pct_cuda_time": 0.02872982388556685, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -8234,8 +8234,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.207, - "pct_cuda_time": 0.019169193813360817, + "cuda_time_us": 5.166, + "pct_cuda_time": 0.018619780478338774, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -8244,8 +8244,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 12.177, - "pct_cuda_time": 0.04482874458715089, + "cuda_time_us": 10.776, + "pct_cuda_time": 0.03883986729279493, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -8254,8 +8254,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 9.012, - "pct_cuda_time": 0.03317702605070246, + "cuda_time_us": 8.812, + "pct_cuda_time": 0.031761034760960365, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -8264,8 +8264,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.408, - "pct_cuda_time": 0.019909160772547594, + "cuda_time_us": 5.247, + "pct_cuda_time": 0.0189117282558737, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -8275,9 +8275,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 431.044, - "cuda_time_us": 146.056, - "pct_cuda_time": 0.5376945979650908, + "cpu_time_us": 456.853, + "cuda_time_us": 149.58200000000002, + "pct_cuda_time": 0.5391374377682676, "trace": "" }, "children": [ @@ -8286,7 +8286,7 @@ "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 9.693, - "pct_cuda_time": 0.03568407828555913, + "pct_cuda_time": 0.03493641737834644, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -8295,8 +8295,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 136.363, - "pct_cuda_time": 0.5020105196795317, + "cuda_time_us": 139.889, + "pct_cuda_time": 0.5042010203899211, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -8306,9 +8306,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.138, - "cuda_time_us": 46.679, - "pct_cuda_time": 0.1718453616312406, + "cpu_time_us": 106.71, + "cuda_time_us": 64.496, + "pct_cuda_time": 0.23246251678879934, "trace": "" }, "children": [ @@ -8316,8 +8316,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 46.679, - "pct_cuda_time": 0.1718453616312406, + "cuda_time_us": 64.496, + "pct_cuda_time": 0.23246251678879934, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -8329,9 +8329,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.407, - "cuda_time_us": 14.02, - "pct_cuda_time": 0.051613615760191786, + "cpu_time_us": 39.129, + "cuda_time_us": 14.741, + "pct_cuda_time": 0.05313089121780717, "trace": "" }, "children": [ @@ -8339,8 +8339,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.02, - "pct_cuda_time": 0.051613615760191786, + "cuda_time_us": 14.741, + "pct_cuda_time": 0.05313089121780717, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8350,18 +8350,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 221.17, - "cuda_time_us": 479.683, - "pct_cuda_time": 1.7659182631024308, + "cpu_time_us": 219.849, + "cuda_time_us": 482.488, + "pct_cuda_time": 1.7390283862626243, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 84.626, - "cuda_time_us": 281.77, - "pct_cuda_time": 1.0373158710948103, + "cpu_time_us": 85.678, + "cuda_time_us": 279.687, + "pct_cuda_time": 1.008074050066809, "trace": "" }, "children": [ @@ -8369,8 +8369,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 281.77, - "pct_cuda_time": 1.0373158710948103, + "cuda_time_us": 279.687, + "pct_cuda_time": 1.008074050066809, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -8380,9 +8380,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.56, - "cuda_time_us": 25.186, - "pct_cuda_time": 0.09272043698546295, + "cpu_time_us": 26.279, + "cuda_time_us": 25.798, + "pct_cuda_time": 0.0929835649980998, "trace": "" }, "children": [ @@ -8390,8 +8390,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 25.186, - "pct_cuda_time": 0.09272043698546295, + "cuda_time_us": 25.798, + "pct_cuda_time": 0.0929835649980998, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -8401,9 +8401,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 92.428, - "cuda_time_us": 172.727, - "pct_cuda_time": 0.6358819550221575, + "cpu_time_us": 89.674, + "cuda_time_us": 177.003, + "pct_cuda_time": 0.6379707711977153, "trace": "" }, "children": [ @@ -8411,8 +8411,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 172.727, - "pct_cuda_time": 0.6358819550221575, + "cuda_time_us": 177.003, + "pct_cuda_time": 0.6379707711977153, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -8426,18 +8426,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1043.828, - "cuda_time_us": 813.671, - "pct_cuda_time": 2.995470923624181, + "cpu_time_us": 1195.798, + "cuda_time_us": 838.279, + "pct_cuda_time": 3.0214035926444724, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 36.654, - "cuda_time_us": 15.022, - "pct_cuda_time": 0.05530240627315273, + "cpu_time_us": 43.265, + "cuda_time_us": 14.661, + "pct_cuda_time": 0.05284254773382205, "trace": "" }, "children": [ @@ -8445,8 +8445,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.022, - "pct_cuda_time": 0.05530240627315273, + "cuda_time_us": 14.661, + "pct_cuda_time": 0.05284254773382205, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8456,18 +8456,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 737.17, - "cuda_time_us": 310.584, - "pct_cuda_time": 1.1433925276222117, + "cpu_time_us": 865.083, + "cuda_time_us": 327.73099999999994, + "pct_cuda_time": 1.1812387293740694, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.065, - "cuda_time_us": 77.725, - "pct_cuda_time": 0.2861389646905069, + "cpu_time_us": 87.861, + "cuda_time_us": 78.687, + "pct_cuda_time": 0.2836110465542088, "trace": "" }, "children": [ @@ -8475,8 +8475,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 77.725, - "pct_cuda_time": 0.2861389646905069, + "cuda_time_us": 78.687, + "pct_cuda_time": 0.2836110465542088, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -8486,9 +8486,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 74.992, - "cuda_time_us": 38.010999999999996, - "pct_cuda_time": 0.1399347466947682, + "cpu_time_us": 81.952, + "cuda_time_us": 37.931, + "pct_cuda_time": 0.13671445863799223, "trace": "" }, "children": [ @@ -8496,8 +8496,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.77, - "pct_cuda_time": 0.02860469289990658, + "cuda_time_us": 7.89, + "pct_cuda_time": 0.02843787610803192, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -8506,8 +8506,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.967, - "pct_cuda_time": 0.018285651175525865, + "cuda_time_us": 5.047, + "pct_cuda_time": 0.018190869545910912, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -8516,8 +8516,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.335, - "pct_cuda_time": 0.041728982499413265, + "cuda_time_us": 10.935, + "pct_cuda_time": 0.039412949967215345, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -8526,8 +8526,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.652, - "pct_cuda_time": 0.031851712093950026, + "cuda_time_us": 9.013, + "pct_cuda_time": 0.032485497764472966, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -8536,8 +8536,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.287, - "pct_cuda_time": 0.01946370802597247, + "cuda_time_us": 5.046, + "pct_cuda_time": 0.018187265252361102, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -8547,9 +8547,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 432.686, - "cuda_time_us": 148.38, - "pct_cuda_time": 0.5462502358414593, + "cpu_time_us": 473.417, + "cuda_time_us": 150.624, + "pct_cuda_time": 0.5428931116471737, "trace": "" }, "children": [ @@ -8557,8 +8557,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.493, - "pct_cuda_time": 0.03494779275403001, + "cuda_time_us": 9.774, + "pct_cuda_time": 0.03522836515588136, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -8567,8 +8567,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 138.887, - "pct_cuda_time": 0.5113024430874292, + "cuda_time_us": 140.85, + "pct_cuda_time": 0.5076647464912922, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -8578,9 +8578,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 110.616, - "cuda_time_us": 46.468, - "pct_cuda_time": 0.17106858039547737, + "cpu_time_us": 185.217, + "cuda_time_us": 60.489, + "pct_cuda_time": 0.21802011253469492, "trace": "" }, "children": [ @@ -8588,8 +8588,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 46.468, - "pct_cuda_time": 0.17106858039547737, + "cuda_time_us": 60.489, + "pct_cuda_time": 0.21802011253469492, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -8601,9 +8601,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.876, - "cuda_time_us": 14.701, - "pct_cuda_time": 0.05412066799504847, + "cpu_time_us": 42.273, + "cuda_time_us": 14.821, + "pct_cuda_time": 0.05341923470179228, "trace": "" }, "children": [ @@ -8611,8 +8611,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.701, - "pct_cuda_time": 0.05412066799504847, + "cuda_time_us": 14.821, + "pct_cuda_time": 0.05341923470179228, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8622,18 +8622,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 210.695, - "cuda_time_us": 473.364, - "pct_cuda_time": 1.742655321733768, + "cpu_time_us": 222.243, + "cuda_time_us": 481.06600000000003, + "pct_cuda_time": 1.733903080834789, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.825, - "cuda_time_us": 275.541, - "pct_cuda_time": 1.0143842582153357, + "cpu_time_us": 90.335, + "cuda_time_us": 281.8, + "pct_cuda_time": 1.015689922337566, "trace": "" }, "children": [ @@ -8641,8 +8641,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 275.541, - "pct_cuda_time": 1.0143842582153357, + "cuda_time_us": 281.8, + "pct_cuda_time": 1.015689922337566, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -8652,9 +8652,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.91, - "cuda_time_us": 25.236, - "pct_cuda_time": 0.09290450836834523, + "cpu_time_us": 27.441, + "cuda_time_us": 25.468, + "pct_cuda_time": 0.09179414812666122, "trace": "" }, "children": [ @@ -8662,8 +8662,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 25.236, - "pct_cuda_time": 0.09290450836834523, + "cuda_time_us": 25.468, + "pct_cuda_time": 0.09179414812666122, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -8673,9 +8673,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 82.303, - "cuda_time_us": 172.587, - "pct_cuda_time": 0.6353665551500871, + "cpu_time_us": 84.687, + "cuda_time_us": 173.798, + "pct_cuda_time": 0.6264190103705617, "trace": "" }, "children": [ @@ -8683,8 +8683,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 172.587, - "pct_cuda_time": 0.6353665551500871, + "cuda_time_us": 173.798, + "pct_cuda_time": 0.6264190103705617, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -8698,18 +8698,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1019.732, - "cuda_time_us": 821.3120000000001, - "pct_cuda_time": 3.023600712356252, + "cpu_time_us": 1095.378, + "cuda_time_us": 830.278, + "pct_cuda_time": 2.992565639952412, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.088, - "cuda_time_us": 14.981, - "pct_cuda_time": 0.05515146773918925, + "cpu_time_us": 34.291, + "cuda_time_us": 15.462, + "pct_cuda_time": 0.05572958686722301, "trace": "" }, "children": [ @@ -8717,8 +8717,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.981, - "pct_cuda_time": 0.05515146773918925, + "cuda_time_us": 15.462, + "pct_cuda_time": 0.05572958686722301, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8728,18 +8728,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 719.064, - "cuda_time_us": 314.29, - "pct_cuda_time": 1.1570358985214464, + "cpu_time_us": 777.593, + "cuda_time_us": 329.715, + "pct_cuda_time": 1.1883896477769005, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 81.411, - "cuda_time_us": 79.738, - "pct_cuda_time": 0.2935496785653476, + "cpu_time_us": 93.721, + "cuda_time_us": 73.86, + "pct_cuda_time": 0.266213121589257, "trace": "" }, "children": [ @@ -8747,8 +8747,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 79.738, - "pct_cuda_time": 0.2935496785653476, + "cuda_time_us": 73.86, + "pct_cuda_time": 0.266213121589257, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -8758,9 +8758,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 74.061, - "cuda_time_us": 39.45399999999999, - "pct_cuda_time": 0.14524704680475084, + "cpu_time_us": 75.764, + "cuda_time_us": 37.932, + "pct_cuda_time": 0.13671806293154207, "trace": "" }, "children": [ @@ -8768,8 +8768,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.851, - "pct_cuda_time": 0.028902888540175878, + "cuda_time_us": 7.73, + "pct_cuda_time": 0.027861189140061696, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -8778,8 +8778,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.407, - "pct_cuda_time": 0.019905479344889947, + "cuda_time_us": 5.047, + "pct_cuda_time": 0.018190869545910912, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -8788,8 +8788,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.777, - "pct_cuda_time": 0.043356173524092635, + "cuda_time_us": 11.296, + "pct_cuda_time": 0.04071409993869817, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -8798,8 +8798,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 9.012, - "pct_cuda_time": 0.03317702605070246, + "cuda_time_us": 8.772, + "pct_cuda_time": 0.03161686301896781, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -8808,8 +8808,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.407, - "pct_cuda_time": 0.019905479344889947, + "cuda_time_us": 5.087, + "pct_cuda_time": 0.01833504128790347, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -8819,9 +8819,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 427.208, - "cuda_time_us": 148.86, - "pct_cuda_time": 0.5480173211171292, + "cpu_time_us": 454.309, + "cuda_time_us": 152.305, + "pct_cuda_time": 0.5489519291044109, "trace": "" }, "children": [ @@ -8829,8 +8829,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 10.014, - "pct_cuda_time": 0.03686581656366338, + "cuda_time_us": 9.853, + "pct_cuda_time": 0.035513104346316667, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -8839,8 +8839,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 138.846, - "pct_cuda_time": 0.5111515045534658, + "cuda_time_us": 142.452, + "pct_cuda_time": 0.5134388247580942, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -8850,9 +8850,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 100.34, - "cuda_time_us": 46.238, - "pct_cuda_time": 0.17022185203421886, + "cpu_time_us": 118.187, + "cuda_time_us": 65.618, + "pct_cuda_time": 0.23650653415169054, "trace": "" }, "children": [ @@ -8860,8 +8860,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 46.238, - "pct_cuda_time": 0.17022185203421886, + "cuda_time_us": 65.618, + "pct_cuda_time": 0.23650653415169054, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -8873,9 +8873,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.977, - "cuda_time_us": 14.821, - "pct_cuda_time": 0.05456243931396595, + "cpu_time_us": 39.6, + "cuda_time_us": 15.061, + "pct_cuda_time": 0.05428426515374763, "trace": "" }, "children": [ @@ -8883,8 +8883,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.821, - "pct_cuda_time": 0.05456243931396595, + "cuda_time_us": 15.061, + "pct_cuda_time": 0.05428426515374763, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8894,18 +8894,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 210.644, - "cuda_time_us": 477.22, - "pct_cuda_time": 1.7568509067816498, + "cpu_time_us": 221.721, + "cuda_time_us": 470.04, + "pct_cuda_time": 1.6941621401545406, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.314, - "cuda_time_us": 275.17, - "pct_cuda_time": 1.0130184485543492, + "cpu_time_us": 86.239, + "cuda_time_us": 274.399, + "pct_cuda_time": 0.9890145457753929, "trace": "" }, "children": [ @@ -8913,8 +8913,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 275.17, - "pct_cuda_time": 1.0130184485543492, + "cuda_time_us": 274.399, + "pct_cuda_time": 0.9890145457753929, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -8924,9 +8924,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.949, - "cuda_time_us": 26.098, - "pct_cuda_time": 0.09607789900923576, + "cpu_time_us": 26.419, + "cuda_time_us": 26.519, + "pct_cuda_time": 0.09558226064751565, "trace": "" }, "children": [ @@ -8934,8 +8934,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 26.098, - "pct_cuda_time": 0.09607789900923576, + "cuda_time_us": 26.519, + "pct_cuda_time": 0.09558226064751565, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -8945,9 +8945,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 83.895, - "cuda_time_us": 175.952, - "pct_cuda_time": 0.6477545592180647, + "cpu_time_us": 91.206, + "cuda_time_us": 169.122, + "pct_cuda_time": 0.6095653337316318, "trace": "" }, "children": [ @@ -8955,8 +8955,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 175.952, - "pct_cuda_time": 0.6477545592180647, + "cuda_time_us": 169.122, + "pct_cuda_time": 0.6095653337316318, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -8970,18 +8970,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1032.001, - "cuda_time_us": 816.875, - "pct_cuda_time": 3.0072662178392777, + "cpu_time_us": 1091.041, + "cuda_time_us": 831.799, + "pct_cuda_time": 2.9980477704416786, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.699, - "cuda_time_us": 14.14, - "pct_cuda_time": 0.052055387079109276, + "cpu_time_us": 33.149, + "cuda_time_us": 15.702, + "pct_cuda_time": 0.05659461731917835, "trace": "" }, "children": [ @@ -8989,8 +8989,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.14, - "pct_cuda_time": 0.052055387079109276, + "cuda_time_us": 15.702, + "pct_cuda_time": 0.05659461731917835, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -9000,18 +9000,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 729.499, - "cuda_time_us": 314.47, - "pct_cuda_time": 1.1576985554998227, + "cpu_time_us": 753.637, + "cuda_time_us": 332.228, + "pct_cuda_time": 1.197447237467583, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 81.101, - "cuda_time_us": 78.476, - "pct_cuda_time": 0.2889037168613988, + "cpu_time_us": 80.761, + "cuda_time_us": 75.652, + "pct_cuda_time": 0.27267201563052357, "trace": "" }, "children": [ @@ -9019,8 +9019,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT8_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW8_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA8_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", "cpu_time_us": 0, - "cuda_time_us": 78.476, - "pct_cuda_time": 0.2889037168613988, + "cuda_time_us": 75.652, + "pct_cuda_time": 0.27267201563052357, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -9030,9 +9030,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 74.541, - "cuda_time_us": 38.092000000000006, - "pct_cuda_time": 0.14023294233503752, + "cpu_time_us": 82.243, + "cuda_time_us": 38.012, + "pct_cuda_time": 0.13700640641552717, "trace": "" }, "children": [ @@ -9040,8 +9040,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 7.69, - "pct_cuda_time": 0.02831017868729493, + "cuda_time_us": 7.891, + "pct_cuda_time": 0.028441480401581735, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False) <- clone(bfloat16[512, 4096], 0) <- contiguous(bfloat16[512, 4096], 0)" }, "children": [] @@ -9051,7 +9051,7 @@ "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 5.247, - "pct_cuda_time": 0.019316450919666644, + "pct_cuda_time": 0.0189117282558737, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False) <- clone(bfloat16[512, 1024], 0) <- contiguous(bfloat16[512, 1024], 0)" }, "children": [] @@ -9060,8 +9060,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.136, - "pct_cuda_time": 0.04099637839554179, + "cuda_time_us": 10.895, + "pct_cuda_time": 0.03926877822522278, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -9070,8 +9070,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 8.532, - "pct_cuda_time": 0.031409940775032556, + "cuda_time_us": 8.812, + "pct_cuda_time": 0.031761034760960365, "trace": "copy_(bfloat16[512, 4096], bfloat16[512, 4096], False)" }, "children": [] @@ -9080,8 +9080,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.487, - "pct_cuda_time": 0.020199993557501596, + "cuda_time_us": 5.167, + "pct_cuda_time": 0.018623384771888584, "trace": "copy_(bfloat16[512, 1024], bfloat16[512, 1024], False)" }, "children": [] @@ -9091,9 +9091,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 435.611, - "cuda_time_us": 150.983, - "pct_cuda_time": 0.5558329920343109, + "cpu_time_us": 442.492, + "cuda_time_us": 150.82399999999998, + "pct_cuda_time": 0.5436139703571364, "trace": "" }, "children": [ @@ -9101,8 +9101,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 10.134, - "pct_cuda_time": 0.03730758788258086, + "cuda_time_us": 9.694, + "pct_cuda_time": 0.034940021671896256, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -9111,8 +9111,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 140.849, - "pct_cuda_time": 0.51852540415173, + "cuda_time_us": 141.13, + "pct_cuda_time": 0.5086739486852402, "trace": "_attention(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], int32[2], int32[2], 512, 512, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None, None)" }, "children": [] @@ -9122,9 +9122,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 101.993, - "cuda_time_us": 46.919, - "pct_cuda_time": 0.1727289042690755, + "cpu_time_us": 112.308, + "cuda_time_us": 67.74, + "pct_cuda_time": 0.2441548450643957, "trace": "" }, "children": [ @@ -9132,8 +9132,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 46.919, - "pct_cuda_time": 0.1727289042690755, + "cuda_time_us": 67.74, + "pct_cuda_time": 0.2441548450643957, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -9145,9 +9145,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.009, - "cuda_time_us": 15.101, - "pct_cuda_time": 0.05559323905810672, + "cpu_time_us": 39.999, + "cuda_time_us": 15.382, + "pct_cuda_time": 0.0554412433832379, "trace": "" }, "children": [ @@ -9155,8 +9155,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.101, - "pct_cuda_time": 0.05559323905810672, + "cuda_time_us": 15.382, + "pct_cuda_time": 0.0554412433832379, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -9166,18 +9166,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 209.793, - "cuda_time_us": 473.16400000000004, - "pct_cuda_time": 1.7419190362022392, + "cpu_time_us": 242.011, + "cuda_time_us": 468.48699999999997, + "pct_cuda_time": 1.6885646722716794, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.955, - "cuda_time_us": 277.994, - "pct_cuda_time": 1.0234148002595407, + "cpu_time_us": 86.349, + "cuda_time_us": 275.531, + "pct_cuda_time": 0.9930946060737824, "trace": "" }, "children": [ @@ -9185,8 +9185,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT192x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT12_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 277.994, - "pct_cuda_time": 1.0234148002595407, + "cuda_time_us": 275.531, + "pct_cuda_time": 0.9930946060737824, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -9196,9 +9196,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.199, - "cuda_time_us": 25.518, - "pct_cuda_time": 0.0939426709678013, + "cpu_time_us": 26.359, + "cuda_time_us": 26.008, + "pct_cuda_time": 0.09374046664356073, "trace": "" }, "children": [ @@ -9206,8 +9206,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 25.518, - "pct_cuda_time": 0.0939426709678013, + "cuda_time_us": 26.008, + "pct_cuda_time": 0.09374046664356073, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] @@ -9217,9 +9217,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 81.642, - "cuda_time_us": 169.652, - "pct_cuda_time": 0.6245615649748971, + "cpu_time_us": 95.653, + "cuda_time_us": 166.948, + "pct_cuda_time": 0.6017295995543362, "trace": "" }, "children": [ @@ -9227,8 +9227,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD7_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO1_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB4_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 169.652, - "pct_cuda_time": 0.6245615649748971, + "cuda_time_us": 166.948, + "pct_cuda_time": 0.6017295995543362, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -9242,9 +9242,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.98, - "cuda_time_us": 14.06, - "pct_cuda_time": 0.05176087286649762, + "cpu_time_us": 40.661, + "cuda_time_us": 14.221, + "pct_cuda_time": 0.051256658571903924, "trace": "" }, "children": [ @@ -9252,8 +9252,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.06, - "pct_cuda_time": 0.05176087286649762, + "cuda_time_us": 14.221, + "pct_cuda_time": 0.051256658571903924, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -9265,9 +9265,9 @@ { "entry": { "name": "LogitsProcessor", - "cpu_time_us": 151.616, - "cuda_time_us": 239.595, - "pct_cuda_time": 0.8820516596336058, + "cpu_time_us": 142.032, + "cuda_time_us": 238.554, + "pct_cuda_time": 0.8598186434823126, "trace": "" }, "children": [ @@ -9275,8 +9275,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, - "cuda_time_us": 4.806, - "pct_cuda_time": 0.017692941322644918, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.01660137609044297, "trace": "index_select(bfloat16[512, 4096], 0, int64[1])" }, "children": [] @@ -9285,8 +9285,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 234.789, - "pct_cuda_time": 0.8643587183109609, + "cuda_time_us": 233.948, + "pct_cuda_time": 0.8432172673918696, "trace": "_rocm_C::wvSplitK(bfloat16[128256, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -9296,9 +9296,9 @@ { "entry": { "name": "Sampler", - "cpu_time_us": 1110.478, - "cuda_time_us": 244.111, - "pct_cuda_time": 0.8986769869355334, + "cpu_time_us": 971.493, + "cuda_time_us": 242.78799999999998, + "pct_cuda_time": 0.8750792223722248, "trace": "" }, "children": [ @@ -9306,8 +9306,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 5.167, - "pct_cuda_time": 0.01902193670705499, + "cuda_time_us": 4.927, + "pct_cuda_time": 0.01775835431993324, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -9316,8 +9316,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 5.127, - "pct_cuda_time": 0.018874679600749168, + "cuda_time_us": 4.966, + "pct_cuda_time": 0.017898921768375986, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -9326,8 +9326,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 5.126, - "pct_cuda_time": 0.018870998173091524, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.017466406542398314, "trace": "copy_(int32[1], int32[1], True) <- _to_copy(int32[1], 3, 0, None, None, True, None) <- to(int32[1], 3, 0, None, None, True, False, None)" }, "children": [] @@ -9336,8 +9336,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 5.167, - "pct_cuda_time": 0.01902193670705499, + "cuda_time_us": 4.967, + "pct_cuda_time": 0.017902526061925796, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -9346,8 +9346,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 5.167, - "pct_cuda_time": 0.01902193670705499, + "cuda_time_us": 5.006, + "pct_cuda_time": 0.018043093510368542, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -9356,8 +9356,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 5.127, - "pct_cuda_time": 0.018874679600749168, + "cuda_time_us": 4.927, + "pct_cuda_time": 0.01775835431993324, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -9366,8 +9366,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 4.365, - "pct_cuda_time": 0.016069431725623196, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.01429102392501224, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -9376,8 +9376,8 @@ "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 2.643, - "pct_cuda_time": 0.009730013299157412, + "cuda_time_us": 2.763, + "pct_cuda_time": 0.009958663078135893, "trace": "copy_(float32[1, 128256], bfloat16[1, 128256], False) <- _to_copy(bfloat16[1, 128256], 6, None, None, None, False, None) <- to(bfloat16[1, 128256], 6, False, False, None)" }, "children": [] @@ -9386,8 +9386,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2})", "cpu_time_us": 0, - "cuda_time_us": 13.9, - "pct_cuda_time": 0.051171844441274324, + "cuda_time_us": 13.218, + "pct_cuda_time": 0.047641552141440544, "trace": "div_(float32[1, 128256], bfloat16[1, 1])" }, "children": [] @@ -9396,8 +9396,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, - "cuda_time_us": 78.236, - "pct_cuda_time": 0.28802017422356385, + "cuda_time_us": 78.757, + "pct_cuda_time": 0.28386334710269584, "trace": "_softmax(float32[1, 128256], -1, False) <- softmax(float32[1, 128256], -1, 6)" }, "children": [] @@ -9406,8 +9406,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, - "cuda_time_us": 44.786, - "pct_cuda_time": 0.1648764190753174, + "cuda_time_us": 44.746, + "pct_cuda_time": 0.1612777191799742, "trace": "_log_softmax(float32[1, 128256], -1, False) <- log_softmax(float32[1, 128256], -1, 6)" }, "children": [] @@ -9416,8 +9416,8 @@ "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.006, - "pct_cuda_time": 0.018429226854174047, + "cuda_time_us": 4.926, + "pct_cuda_time": 0.017754750026383426, "trace": "copy_(int64[1], int32[1], False) <- _to_copy(int32[1], 4, None, None, None, False, None) <- to(int32[1], 4, False, False, None)" }, "children": [] @@ -9426,8 +9426,8 @@ "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 6.89, - "pct_cuda_time": 0.02536503656117842, + "cuda_time_us": 7.17, + "pct_cuda_time": 0.02584278475216589, "trace": "index(float32[1, 128256], None)" }, "children": [] @@ -9436,8 +9436,8 @@ "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4, 4> >(at::native::ReduceOp, unsigned int, long, 4, 4>)", "cpu_time_us": 0, - "cuda_time_us": 52.157, - "pct_cuda_time": 0.19201222233982335, + "cuda_time_us": 52.718, + "pct_cuda_time": 0.19001114735909086, "trace": "argmax(float32[1, 128256], -1, False)" }, "children": [] @@ -9446,8 +9446,8 @@ "entry": { "name": "Memcpy DtoH (Device -> Host)", "cpu_time_us": 0, - "cuda_time_us": 5.247, - "pct_cuda_time": 0.019316450919666644, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.01761057828439087, "trace": "copy_(int64[1], int64[1], False) <- _to_copy(int64[1], 4, 0, None, None, False, None) <- to(int64[1], 4, 0, None, None, False, False, None)" }, "children": [] @@ -9464,24 +9464,24 @@ { "entry": { "name": "LlamaForCausalLM", - "cuda_time_us": 5269.207, - "pct_cuda_time": 92.18287075468547, + "cuda_time_us": 5346.676, + "pct_cuda_time": 92.27128123512604, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", - "cuda_time_us": 3.685, - "pct_cuda_time": 0.06446774224869432, + "cuda_time_us": 4.366, + "pct_cuda_time": 0.07534707804859696, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", - "cuda_time_us": 3.685, - "pct_cuda_time": 0.06446774224869432, + "cuda_time_us": 4.366, + "pct_cuda_time": 0.07534707804859696, "invocations": 1 }, "children": [] @@ -9491,24 +9491,24 @@ { "entry": { "name": "LlamaDecoderLayer", - "cuda_time_us": 5261.076999999999, - "pct_cuda_time": 92.0406393450567, + "cuda_time_us": 5337.704000000001, + "pct_cuda_time": 92.11644523323598, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cuda_time_us": 301.1449999999998, - "pct_cuda_time": 5.2684228601039464, + "cuda_time_us": 303.71599999999984, + "pct_cuda_time": 5.241436820111696, "invocations": 64 }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, unsigned long, unsigned long)", - "cuda_time_us": 4.967, - "pct_cuda_time": 0.08689586858867428, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "invocations": 1 }, "children": [] @@ -9516,8 +9516,8 @@ { "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", - "cuda_time_us": 296.1779999999999, - "pct_cuda_time": 5.181526991515273, + "cuda_time_us": 299.0699999999999, + "pct_cuda_time": 5.161257588638087, "invocations": 63 }, "children": [] @@ -9527,24 +9527,24 @@ { "entry": { "name": "LlamaAttention", - "cuda_time_us": 1550.7610000000002, - "pct_cuda_time": 27.13000283238195, + "cuda_time_us": 1598.959, + "pct_cuda_time": 27.59434002966252, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cuda_time_us": 555.5160000000001, - "pct_cuda_time": 9.718551506926916, + "cuda_time_us": 557.9819999999999, + "pct_cuda_time": 9.629480829984475, "invocations": 32 }, "children": [ { "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", - "cuda_time_us": 555.5160000000001, - "pct_cuda_time": 9.718551506926916, + "cuda_time_us": 557.9819999999999, + "pct_cuda_time": 9.629480829984475, "invocations": 32 }, "children": [] @@ -9554,16 +9554,16 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cuda_time_us": 122.672, - "pct_cuda_time": 2.1461022733057886, + "cuda_time_us": 121.99100000000001, + "pct_cuda_time": 2.1052829588241853, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", - "cuda_time_us": 122.672, - "pct_cuda_time": 2.1461022733057886, + "cuda_time_us": 121.99100000000001, + "pct_cuda_time": 2.1052829588241853, "invocations": 32 }, "children": [] @@ -9573,16 +9573,16 @@ { "entry": { "name": "Attention", - "cuda_time_us": 429.826, - "pct_cuda_time": 7.51965041513902, + "cuda_time_us": 466.71500000000003, + "pct_cuda_time": 8.054423163410657, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", - "cuda_time_us": 90.74300000000001, - "pct_cuda_time": 1.5875159660443068, + "cuda_time_us": 96.06900000000002, + "pct_cuda_time": 1.657929097812795, "invocations": 32 }, "children": [] @@ -9590,8 +9590,8 @@ { "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", - "cuda_time_us": 217.613, - "pct_cuda_time": 3.807060731062448, + "cuda_time_us": 251.14300000000003, + "pct_cuda_time": 4.3341482414930805, "invocations": 32 }, "children": [] @@ -9599,8 +9599,8 @@ { "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", - "cuda_time_us": 121.46999999999998, - "pct_cuda_time": 2.1250737180322656, + "cuda_time_us": 119.503, + "pct_cuda_time": 2.062345824104783, "invocations": 32 }, "children": [] @@ -9610,16 +9610,16 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cuda_time_us": 442.747, - "pct_cuda_time": 7.745698637010221, + "cuda_time_us": 452.2710000000001, + "pct_cuda_time": 7.805153077443198, "invocations": 32 }, "children": [ { "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", - "cuda_time_us": 442.747, - "pct_cuda_time": 7.745698637010221, + "cuda_time_us": 452.2710000000001, + "pct_cuda_time": 7.805153077443198, "invocations": 32 }, "children": [] @@ -9631,24 +9631,24 @@ { "entry": { "name": "LlamaMLP", - "cuda_time_us": 3409.1710000000003, - "pct_cuda_time": 59.64221365257083, + "cuda_time_us": 3435.029000000001, + "pct_cuda_time": 59.280668383461766, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cuda_time_us": 2042.1370000000004, - "pct_cuda_time": 35.72644823677663, + "cuda_time_us": 2058.654, + "pct_cuda_time": 35.52761420363178, "invocations": 32 }, "children": [ { "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", - "cuda_time_us": 2042.1370000000004, - "pct_cuda_time": 35.72644823677663, + "cuda_time_us": 2058.654, + "pct_cuda_time": 35.52761420363178, "invocations": 32 }, "children": [] @@ -9658,16 +9658,16 @@ { "entry": { "name": "SiluAndMul", - "cuda_time_us": 267.48299999999995, - "pct_cuda_time": 4.679518344615333, + "cuda_time_us": 266.529, + "pct_cuda_time": 4.59967507219755, "invocations": 32 }, "children": [ { "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", - "cuda_time_us": 267.48299999999995, - "pct_cuda_time": 4.679518344615333, + "cuda_time_us": 266.529, + "pct_cuda_time": 4.59967507219755, "invocations": 32 }, "children": [] @@ -9677,16 +9677,16 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cuda_time_us": 1099.5510000000002, - "pct_cuda_time": 19.236247071178862, + "cuda_time_us": 1109.8459999999998, + "pct_cuda_time": 19.153379107632418, "invocations": 32 }, "children": [ { "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", - "cuda_time_us": 1099.5510000000002, - "pct_cuda_time": 19.236247071178862, + "cuda_time_us": 1109.8459999999998, + "pct_cuda_time": 19.153379107632418, "invocations": 32 }, "children": [] @@ -9700,16 +9700,16 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cuda_time_us": 4.445, - "pct_cuda_time": 0.0777636673800397, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "invocations": 1 }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", - "cuda_time_us": 4.445, - "pct_cuda_time": 0.0777636673800397, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "invocations": 1 }, "children": [] @@ -9721,16 +9721,16 @@ { "entry": { "name": "LogitsProcessor", - "cuda_time_us": 263.732, - "pct_cuda_time": 4.613895956236813, + "cuda_time_us": 261.13800000000003, + "pct_cuda_time": 4.506638861075245, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", - "cuda_time_us": 2.884, - "pct_cuda_time": 0.05045453694578954, + "cuda_time_us": 3.204, + "pct_cuda_time": 0.05529364133479265, "invocations": 1 }, "children": [] @@ -9738,8 +9738,8 @@ { "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", - "cuda_time_us": 260.848, - "pct_cuda_time": 4.563441419291022, + "cuda_time_us": 257.934, + "pct_cuda_time": 4.4513452197404515, "invocations": 1 }, "children": [] @@ -9749,16 +9749,16 @@ { "entry": { "name": "Sampler", - "cuda_time_us": 183.09799999999998, - "pct_cuda_time": 3.2032332890777298, + "cuda_time_us": 186.704, + "pct_cuda_time": 3.2220799037987287, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy DtoD (Device -> Device)", - "cuda_time_us": 17.419, - "pct_cuda_time": 0.30473910508276975, + "cuda_time_us": 17.619000000000003, + "pct_cuda_time": 0.3040632542689487, "invocations": 7 }, "children": [] @@ -9766,8 +9766,8 @@ { "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 3.244, - "pct_cuda_time": 0.056752606744847875, + "cuda_time_us": 3.444, + "pct_cuda_time": 0.05943548712766101, "invocations": 1 }, "children": [] @@ -9775,8 +9775,8 @@ { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2})", - "cuda_time_us": 9.173, - "pct_cuda_time": 0.1604783174076725, + "cuda_time_us": 10.174, + "pct_cuda_time": 0.17557974623601133, "invocations": 1 }, "children": [] @@ -9784,8 +9784,8 @@ { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", - "cuda_time_us": 64.375, - "pct_cuda_time": 1.126217342539945, + "cuda_time_us": 64.495, + "pct_cuda_time": 1.1130347683793544, "invocations": 1 }, "children": [] @@ -9793,8 +9793,8 @@ { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", - "cuda_time_us": 35.652, - "pct_cuda_time": 0.6237188457667436, + "cuda_time_us": 35.572, + "pct_cuda_time": 0.6138905772663059, "invocations": 1 }, "children": [] @@ -9802,8 +9802,8 @@ { "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 2.202, - "pct_cuda_time": 0.03852319360424014, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.0663558211399119, "invocations": 1 }, "children": [] @@ -9811,8 +9811,8 @@ { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 5.528, - "pct_cuda_time": 0.09671036069220684, + "cuda_time_us": 5.808, + "pct_cuda_time": 0.10023266818741437, "invocations": 1 }, "children": [] @@ -9820,8 +9820,8 @@ { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4, 4> >(at::native::ReduceOp, unsigned int, long, 4, 4>)", - "cuda_time_us": 40.98, - "pct_cuda_time": 0.7169302787928069, + "cuda_time_us": 41.021, + "pct_cuda_time": 0.7079277344552213, "invocations": 1 }, "children": [] @@ -9829,8 +9829,8 @@ { "entry": { "name": "Memcpy DtoH (Device -> Host)", - "cuda_time_us": 4.525, - "pct_cuda_time": 0.07916323844649711, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "invocations": 1 }, "children": [] @@ -9842,18 +9842,18 @@ { "entry": { "name": "LlamaForCausalLM", - "cpu_time_us": 19156.458, - "cuda_time_us": 5269.207, - "pct_cuda_time": 92.18287075468547, + "cpu_time_us": 20324.775, + "cuda_time_us": 5346.676, + "pct_cuda_time": 92.27128123512604, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", - "cpu_time_us": 137.265, - "cuda_time_us": 3.685, - "pct_cuda_time": 0.06446774224869432, + "cpu_time_us": 132.578, + "cuda_time_us": 4.366, + "pct_cuda_time": 0.07534707804859696, "trace": "" }, "children": [ @@ -9861,8 +9861,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, - "cuda_time_us": 3.685, - "pct_cuda_time": 0.06446774224869432, + "cuda_time_us": 4.366, + "pct_cuda_time": 0.07534707804859696, "trace": "index_select(bfloat16[128256, 4096], 0, int64[1]) <- embedding(bfloat16[128256, 4096], int64[1], -1, False, False)" }, "children": [] @@ -9872,18 +9872,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1377.046, - "cuda_time_us": 183.985, - "pct_cuda_time": 3.2187510332770763, + "cpu_time_us": 1321.686, + "cuda_time_us": 192.19600000000003, + "pct_cuda_time": 3.3168591416922, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 182.963, - "cuda_time_us": 4.967, - "pct_cuda_time": 0.08689586858867428, + "cpu_time_us": 168.402, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "" }, "children": [ @@ -9891,8 +9891,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, unsigned long, unsigned long)", "cpu_time_us": 0, - "cuda_time_us": 4.967, - "pct_cuda_time": 0.08689586858867428, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "_C::rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -9902,18 +9902,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 879.223, - "cuda_time_us": 53.515, - "pct_cuda_time": 0.9362255702683518, + "cpu_time_us": 838.143, + "cuda_time_us": 60.163000000000004, + "pct_cuda_time": 1.0382744518180804, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 207.119, - "cuda_time_us": 18.467, - "pct_cuda_time": 0.3230734860533618, + "cpu_time_us": 214.141, + "cuda_time_us": 18.667, + "pct_cuda_time": 0.32214931423114057, "trace": "" }, "children": [ @@ -9921,8 +9921,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 18.467, - "pct_cuda_time": 0.3230734860533618, + "cuda_time_us": 18.667, + "pct_cuda_time": 0.32214931423114057, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -9932,9 +9932,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 92.237, - "cuda_time_us": 4.807, - "pct_cuda_time": 0.08409672645575948, + "cpu_time_us": 83.044, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "" }, "children": [ @@ -9942,8 +9942,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.807, - "pct_cuda_time": 0.08409672645575948, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -9953,9 +9953,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 394.298, - "cuda_time_us": 16.421, - "pct_cuda_time": 0.2872794560287136, + "cpu_time_us": 332.697, + "cuda_time_us": 22.39, + "pct_cuda_time": 0.38639969709301103, "trace": "" }, "children": [ @@ -9963,8 +9963,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.245, - "pct_cuda_time": 0.07426473971389619, + "cuda_time_us": 4.005, + "pct_cuda_time": 0.0691170516684908, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -9973,8 +9973,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.41, - "pct_cuda_time": 0.12963527003061737, + "cuda_time_us": 13.659, + "pct_cuda_time": 0.23572279868662072, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -9983,8 +9983,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.766, - "pct_cuda_time": 0.08337944628420003, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -9994,9 +9994,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 87.511, - "cuda_time_us": 13.82, - "pct_cuda_time": 0.24177590173051713, + "cpu_time_us": 85.337, + "cuda_time_us": 14.38, + "pct_cuda_time": 0.24816559375602942, "trace": "" }, "children": [ @@ -10004,8 +10004,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.82, - "pct_cuda_time": 0.24177590173051713, + "cuda_time_us": 14.38, + "pct_cuda_time": 0.24816559375602942, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10017,9 +10017,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 61.391, - "cuda_time_us": 5.047, - "pct_cuda_time": 0.08829543965513169, + "cpu_time_us": 58.467, + "cuda_time_us": 5.007, + "pct_cuda_time": 0.08640925785371621, "trace": "" }, "children": [ @@ -10027,8 +10027,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.047, - "pct_cuda_time": 0.08829543965513169, + "cuda_time_us": 5.007, + "pct_cuda_time": 0.08640925785371621, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10038,18 +10038,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 215.612, - "cuda_time_us": 120.456, - "pct_cuda_time": 2.1073341547649185, + "cpu_time_us": 211.366, + "cuda_time_us": 122.38000000000001, + "pct_cuda_time": 2.111996200546793, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 72.348, - "cuda_time_us": 72.947, - "pct_cuda_time": 1.2761813823108563, + "cpu_time_us": 72.499, + "cuda_time_us": 75.912, + "pct_cuda_time": 1.3100658242842633, "trace": "" }, "children": [ @@ -10057,8 +10057,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 72.947, - "pct_cuda_time": 1.2761813823108563, + "cuda_time_us": 75.912, + "pct_cuda_time": 1.3100658242842633, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10068,9 +10068,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 50.074, + "cpu_time_us": 45.868, "cuda_time_us": 9.093, - "pct_cuda_time": 0.15907874634121508, + "pct_cuda_time": 0.15692418247730008, "trace": "" }, "children": [ @@ -10079,7 +10079,7 @@ "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, "cuda_time_us": 9.093, - "pct_cuda_time": 0.15907874634121508, + "pct_cuda_time": 0.15692418247730008, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -10089,9 +10089,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 63.545, - "cuda_time_us": 38.416, - "pct_cuda_time": 0.672074026112847, + "cpu_time_us": 60.871, + "cuda_time_us": 37.375, + "pct_cuda_time": 0.6450061937852294, "trace": "" }, "children": [ @@ -10099,8 +10099,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 38.416, - "pct_cuda_time": 0.672074026112847, + "cuda_time_us": 37.375, + "pct_cuda_time": 0.6450061937852294, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -10114,18 +10114,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 626.625, - "cuda_time_us": 183.102, - "pct_cuda_time": 3.2033032676310524, + "cpu_time_us": 663.653, + "cuda_time_us": 182.702, + "pct_cuda_time": 3.1530146252026485, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 30.125, - "cuda_time_us": 4.926, - "pct_cuda_time": 0.08617858841711486, + "cpu_time_us": 32.248, + "cuda_time_us": 5.087, + "pct_cuda_time": 0.08778987311800567, "trace": "" }, "children": [ @@ -10133,8 +10133,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.926, - "pct_cuda_time": 0.08617858841711486, + "cuda_time_us": 5.087, + "pct_cuda_time": 0.08778987311800567, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10144,18 +10144,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 372.476, - "cuda_time_us": 53.513, - "pct_cuda_time": 0.9361905809916905, + "cpu_time_us": 401.039, + "cuda_time_us": 51.831, + "pct_cuda_time": 0.8944833720423339, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 60.21, - "cuda_time_us": 19.348, - "pct_cuda_time": 0.33848626242272395, + "cpu_time_us": 61.592, + "cuda_time_us": 19.628, + "pct_cuda_time": 0.3387339550934176, "trace": "" }, "children": [ @@ -10163,8 +10163,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 19.348, - "pct_cuda_time": 0.33848626242272395, + "cuda_time_us": 19.628, + "pct_cuda_time": 0.3387339550934176, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10174,9 +10174,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 30.846, - "cuda_time_us": 4.125, - "pct_cuda_time": 0.07216538311421007, + "cpu_time_us": 30.566, + "cuda_time_us": 4.406, + "pct_cuda_time": 0.07603738568074168, "trace": "" }, "children": [ @@ -10184,8 +10184,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.125, - "pct_cuda_time": 0.07216538311421007, + "cuda_time_us": 4.406, + "pct_cuda_time": 0.07603738568074168, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -10195,9 +10195,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 178.497, - "cuda_time_us": 15.9, - "pct_cuda_time": 0.2781647494584097, + "cpu_time_us": 196.583, + "cuda_time_us": 13.737, + "pct_cuda_time": 0.2370688985693029, "trace": "" }, "children": [ @@ -10205,8 +10205,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.324, - "pct_cuda_time": 0.05815217781130527, + "cuda_time_us": 3.364, + "pct_cuda_time": 0.058054871863371556, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10215,8 +10215,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 8.371, - "pct_cuda_time": 0.14644761746643697, + "cuda_time_us": 6.729, + "pct_cuda_time": 0.11612700141754673, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10225,8 +10225,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.205, - "pct_cuda_time": 0.07356495418066747, + "cuda_time_us": 3.644, + "pct_cuda_time": 0.06288702528838465, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10236,9 +10236,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 67.13, - "cuda_time_us": 14.14, - "pct_cuda_time": 0.24737418599634678, + "cpu_time_us": 71.728, + "cuda_time_us": 14.06, + "pct_cuda_time": 0.24264313269887158, "trace": "" }, "children": [ @@ -10246,8 +10246,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.14, - "pct_cuda_time": 0.24737418599634678, + "cuda_time_us": 14.06, + "pct_cuda_time": 0.24264313269887158, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10259,9 +10259,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.243, - "cuda_time_us": 4.967, - "pct_cuda_time": 0.08689586858867428, + "cpu_time_us": 37.727, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.09400264180730822, "trace": "" }, "children": [ @@ -10269,8 +10269,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.967, - "pct_cuda_time": 0.08689586858867428, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.09400264180730822, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10280,18 +10280,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 162.412, - "cuda_time_us": 119.696, - "pct_cuda_time": 2.094038229633573, + "cpu_time_us": 162.894, + "cuda_time_us": 120.33699999999999, + "pct_cuda_time": 2.0767387382350004, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 58.037, - "cuda_time_us": 73.629, - "pct_cuda_time": 1.2881127256524056, + "cpu_time_us": 60.0, + "cuda_time_us": 73.869, + "pct_cuda_time": 1.2748083619724713, "trace": "" }, "children": [ @@ -10299,8 +10299,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 73.629, - "pct_cuda_time": 1.2881127256524056, + "cuda_time_us": 73.869, + "pct_cuda_time": 1.2748083619724713, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10310,9 +10310,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.008, - "cuda_time_us": 8.692, - "pct_cuda_time": 0.15206339637059732, + "cpu_time_us": 22.113, + "cuda_time_us": 8.933, + "pct_cuda_time": 0.15416295194872118, "trace": "" }, "children": [ @@ -10320,8 +10320,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.692, - "pct_cuda_time": 0.15206339637059732, + "cuda_time_us": 8.933, + "pct_cuda_time": 0.15416295194872118, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -10331,9 +10331,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 62.914, - "cuda_time_us": 37.375, - "pct_cuda_time": 0.6538621076105701, + "cpu_time_us": 61.371, + "cuda_time_us": 37.535, + "pct_cuda_time": 0.6477674243138083, "trace": "" }, "children": [ @@ -10341,8 +10341,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 37.375, - "pct_cuda_time": 0.6538621076105701, + "cuda_time_us": 37.535, + "pct_cuda_time": 0.6477674243138083, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -10356,18 +10356,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 590.842, - "cuda_time_us": 179.816, - "pct_cuda_time": 3.145815886076315, + "cpu_time_us": 633.217, + "cuda_time_us": 183.94400000000002, + "pct_cuda_time": 3.1744486771807425, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.924, - "cuda_time_us": 4.846, - "pct_cuda_time": 0.08477901735065745, + "cpu_time_us": 30.146, + "cuda_time_us": 4.887, + "pct_cuda_time": 0.08433833495728203, "trace": "" }, "children": [ @@ -10375,8 +10375,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.846, - "pct_cuda_time": 0.08477901735065745, + "cuda_time_us": 4.887, + "pct_cuda_time": 0.08433833495728203, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10386,18 +10386,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 346.227, - "cuda_time_us": 50.268, - "pct_cuda_time": 0.879420479608512, + "cpu_time_us": 374.961, + "cuda_time_us": 53.913000000000004, + "pct_cuda_time": 0.930413884295467, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 57.206, - "cuda_time_us": 19.268, - "pct_cuda_time": 0.33708669135626657, + "cpu_time_us": 58.678, + "cuda_time_us": 18.827, + "pct_cuda_time": 0.3249105447597195, "trace": "" }, "children": [ @@ -10405,8 +10405,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 19.268, - "pct_cuda_time": 0.33708669135626657, + "cuda_time_us": 18.827, + "pct_cuda_time": 0.3249105447597195, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10416,9 +10416,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.288, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.07006602651452395, + "cpu_time_us": 26.7, + "cuda_time_us": 4.325, + "pct_cuda_time": 0.07463951272564863, "trace": "" }, "children": [ @@ -10426,8 +10426,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.07006602651452395, + "cuda_time_us": 4.325, + "pct_cuda_time": 0.07463951272564863, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -10437,9 +10437,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 160.9, - "cuda_time_us": 12.614, - "pct_cuda_time": 0.22067736790367173, + "cpu_time_us": 187.21, + "cuda_time_us": 16.341, + "pct_cuda_time": 0.2820079254219247, "trace": "" }, "children": [ @@ -10447,8 +10447,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.163, - "pct_cuda_time": 0.05533554104005975, + "cuda_time_us": 3.324, + "pct_cuda_time": 0.05736456423122683, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10457,8 +10457,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.408, - "pct_cuda_time": 0.11210564242323834, + "cuda_time_us": 8.892, + "pct_cuda_time": 0.15345538662577282, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10467,8 +10467,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.043, - "pct_cuda_time": 0.05323618444037363, + "cuda_time_us": 4.125, + "pct_cuda_time": 0.07118797456492498, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10478,9 +10478,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 67.24, - "cuda_time_us": 14.381, - "pct_cuda_time": 0.25159039383404974, + "cpu_time_us": 62.724, + "cuda_time_us": 14.42, + "pct_cuda_time": 0.24885590138817412, "trace": "" }, "children": [ @@ -10488,8 +10488,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.381, - "pct_cuda_time": 0.25159039383404974, + "cuda_time_us": 14.42, + "pct_cuda_time": 0.24885590138817412, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10501,9 +10501,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 31.337, - "cuda_time_us": 5.327, - "pct_cuda_time": 0.09319393838773261, + "cpu_time_us": 31.307, + "cuda_time_us": 5.087, + "pct_cuda_time": 0.08778987311800567, "trace": "" }, "children": [ @@ -10511,8 +10511,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.327, - "pct_cuda_time": 0.09319393838773261, + "cuda_time_us": 5.087, + "pct_cuda_time": 0.08778987311800567, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10522,18 +10522,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 155.292, - "cuda_time_us": 119.375, - "pct_cuda_time": 2.0884224507294125, + "cpu_time_us": 163.344, + "cuda_time_us": 120.05700000000002, + "pct_cuda_time": 2.0719065848099882, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 53.921, - "cuda_time_us": 73.589, - "pct_cuda_time": 1.287412940119177, + "cpu_time_us": 57.847, + "cuda_time_us": 74.43, + "pct_cuda_time": 1.2844899265133012, "trace": "" }, "children": [ @@ -10541,8 +10541,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 73.589, - "pct_cuda_time": 1.287412940119177, + "cuda_time_us": 74.43, + "pct_cuda_time": 1.2844899265133012, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10552,9 +10552,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 23.084, - "cuda_time_us": 9.012, - "pct_cuda_time": 0.15766168063642694, + "cpu_time_us": 21.502, + "cuda_time_us": 8.692, + "pct_cuda_time": 0.15000384846504922, "trace": "" }, "children": [ @@ -10562,8 +10562,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 9.012, - "pct_cuda_time": 0.15766168063642694, + "cuda_time_us": 8.692, + "pct_cuda_time": 0.15000384846504922, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -10573,9 +10573,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 61.922, - "cuda_time_us": 36.774, - "pct_cuda_time": 0.6433478299738088, + "cpu_time_us": 65.558, + "cuda_time_us": 36.935, + "pct_cuda_time": 0.6374128098316375, "trace": "" }, "children": [ @@ -10583,8 +10583,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 36.774, - "pct_cuda_time": 0.6433478299738088, + "cuda_time_us": 36.935, + "pct_cuda_time": 0.6374128098316375, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -10598,18 +10598,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 560.507, - "cuda_time_us": 159.189, - "pct_cuda_time": 2.7849539812286026, + "cpu_time_us": 645.295, + "cuda_time_us": 160.387, + "pct_cuda_time": 2.7679092549199087, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 31.367, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.08267966075097134, + "cpu_time_us": 32.529, + "cuda_time_us": 5.047, + "pct_cuda_time": 0.08709956548586095, "trace": "" }, "children": [ @@ -10617,8 +10617,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.08267966075097134, + "cuda_time_us": 5.047, + "pct_cuda_time": 0.08709956548586095, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10628,18 +10628,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 332.226, - "cuda_time_us": 48.227999999999994, - "pct_cuda_time": 0.8437314174138479, + "cpu_time_us": 406.558, + "cuda_time_us": 46.983, + "pct_cuda_time": 0.8108180870263928, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 51.196, - "cuda_time_us": 18.266, - "pct_cuda_time": 0.3195570637488875, + "cpu_time_us": 56.824, + "cuda_time_us": 18.186, + "pct_cuda_time": 0.31384836495460017, "trace": "" }, "children": [ @@ -10647,8 +10647,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 18.266, - "pct_cuda_time": 0.3195570637488875, + "cuda_time_us": 18.186, + "pct_cuda_time": 0.31384836495460017, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10658,9 +10658,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.648, - "cuda_time_us": 4.206, - "pct_cuda_time": 0.0735824488189982, + "cpu_time_us": 26.289, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.06980735930063553, "trace": "" }, "children": [ @@ -10668,8 +10668,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.206, - "pct_cuda_time": 0.0735824488189982, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.06980735930063553, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -10679,9 +10679,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 158.497, - "cuda_time_us": 14.099, - "pct_cuda_time": 0.24665690582478736, + "cpu_time_us": 217.906, + "cuda_time_us": 11.693999999999999, + "pct_cuda_time": 0.20181143625751097, "trace": "" }, "children": [ @@ -10689,8 +10689,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.964, - "pct_cuda_time": 0.051854108012246936, + "cuda_time_us": 2.843, + "pct_cuda_time": 0.04906361495468648, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10699,8 +10699,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.09, - "pct_cuda_time": 0.12403698576478772, + "cuda_time_us": 6.128, + "pct_cuda_time": 0.1057551292445722, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10709,8 +10709,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.045, - "pct_cuda_time": 0.07076581204775266, + "cuda_time_us": 2.723, + "pct_cuda_time": 0.0469926920582523, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10720,9 +10720,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 65.567, - "cuda_time_us": 11.657, - "pct_cuda_time": 0.20393499902117496, + "cpu_time_us": 65.047, + "cuda_time_us": 13.058, + "pct_cuda_time": 0.22535092651364616, "trace": "" }, "children": [ @@ -10730,8 +10730,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 11.657, - "pct_cuda_time": 0.20393499902117496, + "cuda_time_us": 13.058, + "pct_cuda_time": 0.22535092651364616, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10743,9 +10743,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.528, - "cuda_time_us": 4.886, - "pct_cuda_time": 0.08547880288388617, + "cpu_time_us": 30.746, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "" }, "children": [ @@ -10753,8 +10753,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.886, - "pct_cuda_time": 0.08547880288388617, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10764,18 +10764,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 146.318, - "cuda_time_us": 101.349, - "pct_cuda_time": 1.7730641001798972, + "cpu_time_us": 155.042, + "cuda_time_us": 103.67099999999999, + "pct_cuda_time": 1.7891220633019, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 53.319, - "cuda_time_us": 61.851, - "pct_cuda_time": 1.0820608753932137, + "cpu_time_us": 55.713, + "cuda_time_us": 63.494, + "pct_cuda_time": 1.0957598198849328, "trace": "" }, "children": [ @@ -10783,8 +10783,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 61.851, - "pct_cuda_time": 1.0820608753932137, + "cuda_time_us": 63.494, + "pct_cuda_time": 1.0957598198849328, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10794,9 +10794,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 20.31, - "cuda_time_us": 8.412, - "pct_cuda_time": 0.1471648976379964, + "cpu_time_us": 21.953, + "cuda_time_us": 8.491, + "pct_cuda_time": 0.14653505261352195, "trace": "" }, "children": [ @@ -10804,8 +10804,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.412, - "pct_cuda_time": 0.1471648976379964, + "cuda_time_us": 8.491, + "pct_cuda_time": 0.14653505261352195, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -10815,9 +10815,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 56.925, - "cuda_time_us": 31.086, - "pct_cuda_time": 0.5438383271486871, + "cpu_time_us": 60.29, + "cuda_time_us": 31.686, + "pct_cuda_time": 0.5468271908034456, "trace": "" }, "children": [ @@ -10825,8 +10825,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 31.086, - "pct_cuda_time": 0.5438383271486871, + "cuda_time_us": 31.686, + "pct_cuda_time": 0.5468271908034456, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -10840,18 +10840,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 544.964, - "cuda_time_us": 164.204, - "pct_cuda_time": 2.8726895924571516, + "cpu_time_us": 636.632, + "cuda_time_us": 161.45, + "pct_cuda_time": 2.7862541802441547, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.22, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "cpu_time_us": 29.695, + "cuda_time_us": 4.887, + "pct_cuda_time": 0.08433833495728203, "trace": "" }, "children": [ @@ -10859,8 +10859,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "cuda_time_us": 4.887, + "pct_cuda_time": 0.08433833495728203, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10870,18 +10870,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 328.15, - "cuda_time_us": 48.466, - "pct_cuda_time": 0.8478951413365589, + "cpu_time_us": 368.09, + "cuda_time_us": 45.42, + "pct_cuda_time": 0.7838443163003377, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 55.212, - "cuda_time_us": 16.143, - "pct_cuda_time": 0.28241594657277413, + "cpu_time_us": 56.624, + "cuda_time_us": 16.984, + "pct_cuda_time": 0.29310462060865117, "trace": "" }, "children": [ @@ -10889,8 +10889,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.143, - "pct_cuda_time": 0.28241594657277413, + "cuda_time_us": 16.984, + "pct_cuda_time": 0.29310462060865117, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10900,9 +10900,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.538, + "cpu_time_us": 25.749, "cuda_time_us": 3.604, - "pct_cuda_time": 0.0630506765439062, + "pct_cuda_time": 0.06219671765623992, "trace": "" }, "children": [ @@ -10911,7 +10911,7 @@ "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.604, - "pct_cuda_time": 0.0630506765439062, + "pct_cuda_time": 0.06219671765623992, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -10921,9 +10921,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 151.677, - "cuda_time_us": 15.54, - "pct_cuda_time": 0.27186667965935135, + "cpu_time_us": 187.6, + "cuda_time_us": 12.174, + "pct_cuda_time": 0.2100951278432477, "trace": "" }, "children": [ @@ -10931,8 +10931,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.923, - "pct_cuda_time": 0.05113682784068752, + "cuda_time_us": 3.003, + "pct_cuda_time": 0.05182484548326539, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10941,8 +10941,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 8.652, - "pct_cuda_time": 0.1513636108373686, + "cuda_time_us": 6.328, + "pct_cuda_time": 0.10920666740529585, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10951,8 +10951,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06936624098129525, + "cuda_time_us": 2.843, + "pct_cuda_time": 0.04906361495468648, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10962,9 +10962,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 66.569, - "cuda_time_us": 13.179, - "pct_cuda_time": 0.23056183856052714, + "cpu_time_us": 64.697, + "cuda_time_us": 12.658, + "pct_cuda_time": 0.2184478501921989, "trace": "" }, "children": [ @@ -10972,8 +10972,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.179, - "pct_cuda_time": 0.23056183856052714, + "cuda_time_us": 12.658, + "pct_cuda_time": 0.2184478501921989, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10985,9 +10985,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 30.686, - "cuda_time_us": 4.806, - "pct_cuda_time": 0.08407923181742875, + "cpu_time_us": 29.334, + "cuda_time_us": 4.727, + "pct_cuda_time": 0.08157710442870314, "trace": "" }, "children": [ @@ -10995,8 +10995,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.806, - "pct_cuda_time": 0.08407923181742875, + "cuda_time_us": 4.727, + "pct_cuda_time": 0.08157710442870314, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11006,18 +11006,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 138.286, - "cuda_time_us": 106.246, - "pct_cuda_time": 1.858735344085421, + "cpu_time_us": 188.812, + "cuda_time_us": 106.416, + "pct_cuda_time": 1.8364944245578319, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 50.665, - "cuda_time_us": 63.494, - "pct_cuda_time": 1.1108045661705828, + "cpu_time_us": 54.081, + "cuda_time_us": 62.552, + "pct_cuda_time": 1.0795030751479242, "trace": "" }, "children": [ @@ -11025,8 +11025,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 63.494, - "pct_cuda_time": 1.1108045661705828, + "cuda_time_us": 62.552, + "pct_cuda_time": 1.0795030751479242, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11036,9 +11036,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 17.386, - "cuda_time_us": 8.452, - "pct_cuda_time": 0.1478646831712251, + "cpu_time_us": 23.295, + "cuda_time_us": 8.572, + "pct_cuda_time": 0.147932925568615, "trace": "" }, "children": [ @@ -11046,8 +11046,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.452, - "pct_cuda_time": 0.1478646831712251, + "cuda_time_us": 8.572, + "pct_cuda_time": 0.147932925568615, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -11057,9 +11057,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 55.143, - "cuda_time_us": 34.3, - "pct_cuda_time": 0.6000660947436134, + "cpu_time_us": 92.468, + "cuda_time_us": 35.292, + "pct_cuda_time": 0.6090584238412928, "trace": "" }, "children": [ @@ -11067,8 +11067,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 34.3, - "pct_cuda_time": 0.6000660947436134, + "cuda_time_us": 35.292, + "pct_cuda_time": 0.6090584238412928, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -11082,18 +11082,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 632.545, - "cuda_time_us": 160.769, - "pct_cuda_time": 2.812595509791137, + "cpu_time_us": 617.022, + "cuda_time_us": 165.02599999999998, + "pct_cuda_time": 2.847967682557893, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 57.246, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "cpu_time_us": 28.412, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "" }, "children": [ @@ -11101,8 +11101,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11112,18 +11112,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 345.987, - "cuda_time_us": 45.722, - "pct_cuda_time": 0.7998898537570697, + "cpu_time_us": 369.833, + "cuda_time_us": 49.788999999999994, + "pct_cuda_time": 0.8592431674213453, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 53.139, - "cuda_time_us": 16.664, - "pct_cuda_time": 0.29153065314307797, + "cpu_time_us": 56.665, + "cuda_time_us": 18.677, + "pct_cuda_time": 0.3223218911391767, "trace": "" }, "children": [ @@ -11131,8 +11131,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.664, - "pct_cuda_time": 0.29153065314307797, + "cuda_time_us": 18.677, + "pct_cuda_time": 0.3223218911391767, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11142,9 +11142,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 28.713, - "cuda_time_us": 3.924, - "pct_cuda_time": 0.06864896080973584, + "cpu_time_us": 27.111, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.0663558211399119, "trace": "" }, "children": [ @@ -11152,8 +11152,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.924, - "pct_cuda_time": 0.06864896080973584, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.0663558211399119, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -11163,9 +11163,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 169.553, - "cuda_time_us": 11.414, - "pct_cuda_time": 0.1996838019068106, + "cpu_time_us": 178.687, + "cuda_time_us": 12.335999999999999, + "pct_cuda_time": 0.21289087375343385, "trace": "" }, "children": [ @@ -11173,8 +11173,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.683, - "pct_cuda_time": 0.046938114641315296, + "cuda_time_us": 2.603, + "pct_cuda_time": 0.04492176916181812, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11183,8 +11183,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.048, - "pct_cuda_time": 0.10580757262418002, + "cuda_time_us": 6.93, + "pct_cuda_time": 0.11959579726907396, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11193,8 +11193,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.683, - "pct_cuda_time": 0.046938114641315296, + "cuda_time_us": 2.803, + "pct_cuda_time": 0.04837330732254175, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11204,9 +11204,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 64.055, - "cuda_time_us": 13.72, - "pct_cuda_time": 0.24002643789744538, + "cpu_time_us": 72.148, + "cuda_time_us": 14.931, + "pct_cuda_time": 0.257674581388823, "trace": "" }, "children": [ @@ -11214,8 +11214,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.72, - "pct_cuda_time": 0.24002643789744538, + "cuda_time_us": 14.931, + "pct_cuda_time": 0.257674581388823, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11227,9 +11227,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.609, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cpu_time_us": 32.869, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "" }, "children": [ @@ -11237,8 +11237,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11248,18 +11248,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 163.534, - "cuda_time_us": 105.715, - "pct_cuda_time": 1.8494456911318105, + "cpu_time_us": 152.998, + "cuda_time_us": 105.865, + "pct_cuda_time": 1.8269854369250382, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 60.801, - "cuda_time_us": 63.544, - "pct_cuda_time": 1.1116792980871186, + "cpu_time_us": 55.523, + "cuda_time_us": 63.224, + "pct_cuda_time": 1.0911002433679557, "trace": "" }, "children": [ @@ -11267,8 +11267,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 63.544, - "pct_cuda_time": 1.1116792980871186, + "cuda_time_us": 63.224, + "pct_cuda_time": 1.0911002433679557, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11278,9 +11278,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 23.025, - "cuda_time_us": 8.251, - "pct_cuda_time": 0.14434826086675084, + "cpu_time_us": 24.326, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.14101259155636414, "trace": "" }, "children": [ @@ -11288,8 +11288,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.251, - "pct_cuda_time": 0.14434826086675084, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.14101259155636414, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -11299,9 +11299,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 61.022, - "cuda_time_us": 33.92, - "pct_cuda_time": 0.5934181321779408, + "cpu_time_us": 56.204, + "cuda_time_us": 34.47, + "pct_cuda_time": 0.5948726020007186, "trace": "" }, "children": [ @@ -11309,8 +11309,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.92, - "pct_cuda_time": 0.5934181321779408, + "cuda_time_us": 34.47, + "pct_cuda_time": 0.5948726020007186, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -11324,18 +11324,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 699.624, - "cuda_time_us": 162.45000000000002, - "pct_cuda_time": 2.842003996825073, + "cpu_time_us": 637.453, + "cuda_time_us": 166.15699999999998, + "pct_cuda_time": 2.8674861308567854, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 31.257, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.08267966075097134, + "cpu_time_us": 31.237, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "" }, "children": [ @@ -11343,8 +11343,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.08267966075097134, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11354,18 +11354,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 419.456, - "cuda_time_us": 47.574, - "pct_cuda_time": 0.8322899239455587, + "cpu_time_us": 400.579, + "cuda_time_us": 50.189, + "pct_cuda_time": 0.8661462437427927, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 104.516, - "cuda_time_us": 20.129, - "pct_cuda_time": 0.35214957495901444, + "cpu_time_us": 89.624, + "cuda_time_us": 17.585, + "pct_cuda_time": 0.3034764927816257, "trace": "" }, "children": [ @@ -11373,8 +11373,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 20.129, - "pct_cuda_time": 0.35214957495901444, + "cuda_time_us": 17.585, + "pct_cuda_time": 0.3034764927816257, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11384,9 +11384,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.879, - "cuda_time_us": 3.724, - "pct_cuda_time": 0.06515003314359231, + "cpu_time_us": 33.31, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06497520587562244, "trace": "" }, "children": [ @@ -11394,8 +11394,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.724, - "pct_cuda_time": 0.06515003314359231, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06497520587562244, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -11405,9 +11405,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 190.044, - "cuda_time_us": 11.494, - "pct_cuda_time": 0.201083372973268, + "cpu_time_us": 172.978, + "cuda_time_us": 14.097999999999999, + "pct_cuda_time": 0.24329892494940908, "trace": "" }, "children": [ @@ -11415,8 +11415,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.923, - "pct_cuda_time": 0.05113682784068752, + "cuda_time_us": 3.044, + "pct_cuda_time": 0.052532410806213736, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11425,8 +11425,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.167, - "pct_cuda_time": 0.0903947962548178, + "cuda_time_us": 7.29, + "pct_cuda_time": 0.12580856595837653, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11435,8 +11435,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.404, - "pct_cuda_time": 0.059551748877762684, + "cuda_time_us": 3.764, + "pct_cuda_time": 0.06495794818481881, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11446,9 +11446,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 63.735, - "cuda_time_us": 12.227, - "pct_cuda_time": 0.213906942869684, + "cpu_time_us": 65.939, + "cuda_time_us": 14.741, + "pct_cuda_time": 0.25439562013613554, "trace": "" }, "children": [ @@ -11456,8 +11456,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 12.227, - "pct_cuda_time": 0.213906942869684, + "cuda_time_us": 14.741, + "pct_cuda_time": 0.25439562013613554, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11469,9 +11469,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.975, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "cpu_time_us": 32.799, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "" }, "children": [ @@ -11479,8 +11479,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11490,18 +11490,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 197.845, - "cuda_time_us": 105.54400000000001, - "pct_cuda_time": 1.8464541079772578, + "cpu_time_us": 153.189, + "cuda_time_us": 106.556, + "pct_cuda_time": 1.8389105012703386, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 86.038, - "cuda_time_us": 62.132, - "pct_cuda_time": 1.0869768687641455, + "cpu_time_us": 55.934, + "cuda_time_us": 63.755, + "pct_cuda_time": 1.100264077184677, "trace": "" }, "children": [ @@ -11509,8 +11509,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.132, - "pct_cuda_time": 1.0869768687641455, + "cuda_time_us": 63.755, + "pct_cuda_time": 1.100264077184677, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11520,9 +11520,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 27.18, - "cuda_time_us": 8.291, - "pct_cuda_time": 0.14504804639997956, + "cpu_time_us": 24.296, + "cuda_time_us": 8.331, + "pct_cuda_time": 0.14377382208494305, "trace": "" }, "children": [ @@ -11530,8 +11530,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.291, - "pct_cuda_time": 0.14504804639997956, + "cuda_time_us": 8.331, + "pct_cuda_time": 0.14377382208494305, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -11541,9 +11541,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 64.866, - "cuda_time_us": 35.121, - "pct_cuda_time": 0.6144291928131326, + "cpu_time_us": 54.391, + "cuda_time_us": 34.47, + "pct_cuda_time": 0.5948726020007186, "trace": "" }, "children": [ @@ -11551,8 +11551,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.121, - "pct_cuda_time": 0.6144291928131326, + "cuda_time_us": 34.47, + "pct_cuda_time": 0.5948726020007186, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -11566,18 +11566,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 605.083, - "cuda_time_us": 162.933, - "pct_cuda_time": 2.8504539071388093, + "cpu_time_us": 596.942, + "cuda_time_us": 165.537, + "pct_cuda_time": 2.856786362558543, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 31.727, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07988051861805652, + "cpu_time_us": 29.925, + "cuda_time_us": 4.767, + "pct_cuda_time": 0.08226741206084787, "trace": "" }, "children": [ @@ -11585,8 +11585,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07988051861805652, + "cuda_time_us": 4.767, + "pct_cuda_time": 0.08226741206084787, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11596,18 +11596,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 360.818, - "cuda_time_us": 47.195, - "pct_cuda_time": 0.8256594560182167, + "cpu_time_us": 353.749, + "cuda_time_us": 50.720000000000006, + "pct_cuda_time": 0.8753100775595142, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 57.897, - "cuda_time_us": 16.494, - "pct_cuda_time": 0.28855656462685597, + "cpu_time_us": 57.426, + "cuda_time_us": 17.725, + "pct_cuda_time": 0.30589256949413224, "trace": "" }, "children": [ @@ -11615,8 +11615,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.494, - "pct_cuda_time": 0.28855656462685597, + "cuda_time_us": 17.725, + "pct_cuda_time": 0.30589256949413224, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11626,9 +11626,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 28.883, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.06796666991483784, + "cpu_time_us": 26.65, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06566551350776717, "trace": "" }, "children": [ @@ -11636,8 +11636,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.06796666991483784, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06566551350776717, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -11647,9 +11647,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 177.505, - "cuda_time_us": 12.254999999999999, - "pct_cuda_time": 0.21439679274294407, + "cpu_time_us": 173.539, + "cuda_time_us": 14.459, + "pct_cuda_time": 0.24952895132951525, "trace": "" }, "children": [ @@ -11657,8 +11657,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.723, - "pct_cuda_time": 0.047637900174544, + "cuda_time_us": 2.603, + "pct_cuda_time": 0.04492176916181812, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11667,8 +11667,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.128, - "pct_cuda_time": 0.10720714369063741, + "cuda_time_us": 8.011, + "pct_cuda_time": 0.1382513610277852, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11677,8 +11677,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.404, - "pct_cuda_time": 0.059551748877762684, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.0663558211399119, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11688,9 +11688,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 59.028, - "cuda_time_us": 14.561, - "pct_cuda_time": 0.25473942873357885, + "cpu_time_us": 60.711, + "cuda_time_us": 14.731, + "pct_cuda_time": 0.2542230432280994, "trace": "" }, "children": [ @@ -11698,8 +11698,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.561, - "pct_cuda_time": 0.25473942873357885, + "cuda_time_us": 14.731, + "pct_cuda_time": 0.2542230432280994, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11711,9 +11711,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.879, - "cuda_time_us": 4.886, - "pct_cuda_time": 0.08547880288388617, + "cpu_time_us": 30.295, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "" }, "children": [ @@ -11721,8 +11721,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.886, - "pct_cuda_time": 0.08547880288388617, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11732,18 +11732,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 157.325, - "cuda_time_us": 106.286, - "pct_cuda_time": 1.85943512961865, + "cpu_time_us": 158.136, + "cuda_time_us": 105.404, + "pct_cuda_time": 1.8190296414645706, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 56.604, - "cuda_time_us": 63.283, - "pct_cuda_time": 1.1071131974828015, + "cpu_time_us": 55.773, + "cuda_time_us": 62.052, + "pct_cuda_time": 1.0708742297461151, "trace": "" }, "children": [ @@ -11751,8 +11751,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 63.283, - "pct_cuda_time": 1.1071131974828015, + "cuda_time_us": 62.052, + "pct_cuda_time": 1.0708742297461151, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11762,9 +11762,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 24.366, - "cuda_time_us": 8.252, - "pct_cuda_time": 0.1443657555050816, + "cpu_time_us": 23.275, + "cuda_time_us": 8.091, + "pct_cuda_time": 0.1396319762920747, "trace": "" }, "children": [ @@ -11772,8 +11772,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.252, - "pct_cuda_time": 0.1443657555050816, + "cuda_time_us": 8.091, + "pct_cuda_time": 0.1396319762920747, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -11783,9 +11783,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 59.109, - "cuda_time_us": 34.751, - "pct_cuda_time": 0.607956176630767, + "cpu_time_us": 61.311, + "cuda_time_us": 35.261, + "pct_cuda_time": 0.6085234354263805, "trace": "" }, "children": [ @@ -11793,8 +11793,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 34.751, - "pct_cuda_time": 0.607956176630767, + "cuda_time_us": 35.261, + "pct_cuda_time": 0.6085234354263805, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -11808,18 +11808,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 588.809, - "cuda_time_us": 166.067, - "pct_cuda_time": 2.9052821036672785, + "cpu_time_us": 598.765, + "cuda_time_us": 164.507, + "pct_cuda_time": 2.839010941030816, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.272, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.07918073308482781, + "cpu_time_us": 29.434, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "" }, "children": [ @@ -11827,8 +11827,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.07918073308482781, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11838,18 +11838,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 351.775, - "cuda_time_us": 47.985, - "pct_cuda_time": 0.8394802202994835, + "cpu_time_us": 361.781, + "cuda_time_us": 48.166999999999994, + "pct_cuda_time": 0.8312511929378766, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 52.118, - "cuda_time_us": 16.854, - "pct_cuda_time": 0.29485463442591425, + "cpu_time_us": 54.351, + "cuda_time_us": 16.724, + "pct_cuda_time": 0.28861762099971044, "trace": "" }, "children": [ @@ -11857,8 +11857,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.854, - "pct_cuda_time": 0.29485463442591425, + "cuda_time_us": 16.724, + "pct_cuda_time": 0.28861762099971044, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11868,9 +11868,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 28.072, - "cuda_time_us": 3.524, - "pct_cuda_time": 0.061651105477448793, + "cpu_time_us": 26.75, + "cuda_time_us": 3.525, + "pct_cuda_time": 0.06083336008275408, "trace": "" }, "children": [ @@ -11878,8 +11878,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.524, - "pct_cuda_time": 0.061651105477448793, + "cuda_time_us": 3.525, + "pct_cuda_time": 0.06083336008275408, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -11889,9 +11889,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 176.193, - "cuda_time_us": 13.978, - "pct_cuda_time": 0.2445400545867705, + "cpu_time_us": 183.284, + "cuda_time_us": 13.417, + "pct_cuda_time": 0.2315464375121451, "trace": "" }, "children": [ @@ -11899,8 +11899,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.603, - "pct_cuda_time": 0.0455385435748579, + "cuda_time_us": 2.643, + "pct_cuda_time": 0.04561207679396284, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11909,8 +11909,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.49, - "pct_cuda_time": 0.13103484109707478, + "cuda_time_us": 6.569, + "pct_cuda_time": 0.11336577088896782, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11919,8 +11919,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.06796666991483784, + "cuda_time_us": 4.205, + "pct_cuda_time": 0.07256858982921445, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11930,9 +11930,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 62.253, - "cuda_time_us": 13.629, - "pct_cuda_time": 0.23843442580935006, + "cpu_time_us": 60.771, + "cuda_time_us": 14.501, + "pct_cuda_time": 0.2502537743432672, "trace": "" }, "children": [ @@ -11940,8 +11940,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.629, - "pct_cuda_time": 0.23843442580935006, + "cuda_time_us": 14.501, + "pct_cuda_time": 0.2502537743432672, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11953,9 +11953,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 36.003, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.08267966075097134, + "cpu_time_us": 33.981, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.07810830857717588, "trace": "" }, "children": [ @@ -11963,8 +11963,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.08267966075097134, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.07810830857717588, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11974,18 +11974,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 153.409, - "cuda_time_us": 108.83, - "pct_cuda_time": 1.9039414895319955, + "cpu_time_us": 153.45, + "cuda_time_us": 107.208, + "pct_cuda_time": 1.8501625156742978, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 55.223, - "cuda_time_us": 65.337, - "pct_cuda_time": 1.1430471846140953, + "cpu_time_us": 53.981, + "cuda_time_us": 63.283, + "pct_cuda_time": 1.0921184471253693, "trace": "" }, "children": [ @@ -11993,8 +11993,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 65.337, - "pct_cuda_time": 1.1430471846140953, + "cuda_time_us": 63.283, + "pct_cuda_time": 1.0921184471253693, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12004,9 +12004,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 21.963, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14154911873383605, + "cpu_time_us": 24.406, + "cuda_time_us": 8.172, + "pct_cuda_time": 0.14102984924716777, "trace": "" }, "children": [ @@ -12014,8 +12014,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14154911873383605, + "cuda_time_us": 8.172, + "pct_cuda_time": 0.14102984924716777, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -12025,9 +12025,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 57.446, - "cuda_time_us": 35.402, - "pct_cuda_time": 0.6193451861840642, + "cpu_time_us": 56.304, + "cuda_time_us": 35.753, + "pct_cuda_time": 0.6170142193017607, "trace": "" }, "children": [ @@ -12035,8 +12035,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.402, - "pct_cuda_time": 0.6193451861840642, + "cuda_time_us": 35.753, + "pct_cuda_time": 0.6170142193017607, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -12050,18 +12050,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 579.055, - "cuda_time_us": 159.728, - "pct_cuda_time": 2.79438359128886, + "cpu_time_us": 599.276, + "cuda_time_us": 166.248, + "pct_cuda_time": 2.869056580719915, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.775, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.07918073308482781, + "cpu_time_us": 31.377, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "" }, "children": [ @@ -12069,8 +12069,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.07918073308482781, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12080,18 +12080,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 352.175, - "cuda_time_us": 44.781, - "pct_cuda_time": 0.7834273990878646, + "cpu_time_us": 353.198, + "cuda_time_us": 50.66, + "pct_cuda_time": 0.874274616111297, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 53.119, - "cuda_time_us": 16.764, - "pct_cuda_time": 0.2932801169761497, + "cpu_time_us": 54.341, + "cuda_time_us": 17.575, + "pct_cuda_time": 0.30330391587358946, "trace": "" }, "children": [ @@ -12099,8 +12099,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.764, - "pct_cuda_time": 0.2932801169761497, + "cuda_time_us": 17.575, + "pct_cuda_time": 0.30330391587358946, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12110,9 +12110,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 22.784, - "cuda_time_us": 3.884, - "pct_cuda_time": 0.06794917527650712, + "cpu_time_us": 29.204, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06566551350776717, "trace": "" }, "children": [ @@ -12120,8 +12120,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.884, - "pct_cuda_time": 0.06794917527650712, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06566551350776717, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -12131,9 +12131,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 183.213, - "cuda_time_us": 10.934, - "pct_cuda_time": 0.19128637550806615, + "cpu_time_us": 179.508, + "cuda_time_us": 15.020000000000001, + "pct_cuda_time": 0.25921051587034505, "trace": "" }, "children": [ @@ -12141,8 +12141,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.723, - "pct_cuda_time": 0.047637900174544, + "cuda_time_us": 2.563, + "pct_cuda_time": 0.044231461529673395, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12151,8 +12151,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.648, - "pct_cuda_time": 0.09880971729189297, + "cuda_time_us": 8.612, + "pct_cuda_time": 0.14862323320075976, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12161,8 +12161,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.563, - "pct_cuda_time": 0.04483875804162919, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.0663558211399119, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12172,9 +12172,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 60.29, - "cuda_time_us": 13.199, - "pct_cuda_time": 0.23091173132714152, + "cpu_time_us": 62.563, + "cuda_time_us": 14.26, + "pct_cuda_time": 0.24609467085959522, "trace": "" }, "children": [ @@ -12182,8 +12182,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.199, - "pct_cuda_time": 0.23091173132714152, + "cuda_time_us": 14.26, + "pct_cuda_time": 0.24609467085959522, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12195,9 +12195,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.394, + "cpu_time_us": 42.234, "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "pct_cuda_time": 0.08086953910575478, "trace": "" }, "children": [ @@ -12206,7 +12206,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "pct_cuda_time": 0.08086953910575478, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12216,18 +12216,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 147.34, - "cuda_time_us": 105.735, - "pct_cuda_time": 1.8497955838984246, + "cpu_time_us": 153.159, + "cuda_time_us": 106.296, + "pct_cuda_time": 1.8344235016613981, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 53.439, - "cuda_time_us": 62.292, - "pct_cuda_time": 1.0897760108970602, + "cpu_time_us": 57.586, + "cuda_time_us": 64.425, + "pct_cuda_time": 1.111826730023101, "trace": "" }, "children": [ @@ -12235,8 +12235,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.292, - "pct_cuda_time": 1.0897760108970602, + "cuda_time_us": 64.425, + "pct_cuda_time": 1.111826730023101, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12246,9 +12246,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 20.701, - "cuda_time_us": 8.211, - "pct_cuda_time": 0.14364847533352215, + "cpu_time_us": 21.842, + "cuda_time_us": 8.332, + "pct_cuda_time": 0.14379107977574668, "trace": "" }, "children": [ @@ -12256,8 +12256,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.211, - "pct_cuda_time": 0.14364847533352215, + "cuda_time_us": 8.332, + "pct_cuda_time": 0.14379107977574668, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -12267,9 +12267,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 55.433, - "cuda_time_us": 35.232, - "pct_cuda_time": 0.6163710976678423, + "cpu_time_us": 56.144, + "cuda_time_us": 33.539, + "pct_cuda_time": 0.57880569186255, "trace": "" }, "children": [ @@ -12277,8 +12277,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.232, - "pct_cuda_time": 0.6163710976678423, + "cuda_time_us": 33.539, + "pct_cuda_time": 0.57880569186255, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -12292,18 +12292,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 581.738, - "cuda_time_us": 165.686, - "pct_cuda_time": 2.8986166464632754, + "cpu_time_us": 611.354, + "cuda_time_us": 165.505, + "pct_cuda_time": 2.856234116452826, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.039, + "cpu_time_us": 27.711, "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "pct_cuda_time": 0.08017923147361006, "trace": "" }, "children": [ @@ -12312,7 +12312,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "pct_cuda_time": 0.08017923147361006, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12322,18 +12322,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 357.684, - "cuda_time_us": 47.955000000000005, - "pct_cuda_time": 0.8389553811495621, + "cpu_time_us": 374.489, + "cuda_time_us": 48.515, + "pct_cuda_time": 0.837256869337536, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 51.637, - "cuda_time_us": 17.515, - "pct_cuda_time": 0.3064185903625187, + "cpu_time_us": 56.735, + "cuda_time_us": 17.315, + "pct_cuda_time": 0.2988169162646488, "trace": "" }, "children": [ @@ -12341,8 +12341,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.515, - "pct_cuda_time": 0.3064185903625187, + "cuda_time_us": 17.315, + "pct_cuda_time": 0.2988169162646488, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12352,9 +12352,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 26.329, + "cpu_time_us": 26.5, "cuda_time_us": 3.604, - "pct_cuda_time": 0.0630506765439062, + "pct_cuda_time": 0.06219671765623992, "trace": "" }, "children": [ @@ -12363,7 +12363,7 @@ "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.604, - "pct_cuda_time": 0.0630506765439062, + "pct_cuda_time": 0.06219671765623992, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -12373,9 +12373,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 188.702, - "cuda_time_us": 14.459, - "pct_cuda_time": 0.25295497562384567, + "cpu_time_us": 192.347, + "cuda_time_us": 15.86, + "pct_cuda_time": 0.2737069761453843, "trace": "" }, "children": [ @@ -12384,7 +12384,7 @@ "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.723, - "pct_cuda_time": 0.047637900174544, + "pct_cuda_time": 0.0469926920582523, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12393,8 +12393,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.771, - "pct_cuda_time": 0.1359508344680064, + "cuda_time_us": 9.332, + "pct_cuda_time": 0.16104877057936484, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12403,8 +12403,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06936624098129525, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06566551350776717, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12414,9 +12414,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 57.616, - "cuda_time_us": 12.377, - "pct_cuda_time": 0.21653113861929166, + "cpu_time_us": 64.176, + "cuda_time_us": 11.736, + "pct_cuda_time": 0.20253625927126295, "trace": "" }, "children": [ @@ -12424,8 +12424,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 12.377, - "pct_cuda_time": 0.21653113861929166, + "cuda_time_us": 11.736, + "pct_cuda_time": 0.20253625927126295, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12437,9 +12437,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.911, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "cpu_time_us": 31.808, + "cuda_time_us": 5.127, + "pct_cuda_time": 0.08848018075015039, "trace": "" }, "children": [ @@ -12447,8 +12447,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "cuda_time_us": 5.127, + "pct_cuda_time": 0.08848018075015039, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12458,18 +12458,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 151.095, - "cuda_time_us": 108.399, - "pct_cuda_time": 1.8964013004114564, + "cpu_time_us": 157.936, + "cuda_time_us": 107.217, + "pct_cuda_time": 1.8503178348915301, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 55.463, - "cuda_time_us": 64.345, - "pct_cuda_time": 1.1256925033900234, + "cpu_time_us": 57.086, + "cuda_time_us": 63.804, + "pct_cuda_time": 1.1011097040340543, "trace": "" }, "children": [ @@ -12477,8 +12477,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 64.345, - "pct_cuda_time": 1.1256925033900234, + "cuda_time_us": 63.804, + "pct_cuda_time": 1.1011097040340543, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12488,9 +12488,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 21.382, - "cuda_time_us": 8.251, - "pct_cuda_time": 0.14434826086675084, + "cpu_time_us": 23.164, + "cuda_time_us": 8.332, + "pct_cuda_time": 0.14379107977574668, "trace": "" }, "children": [ @@ -12498,8 +12498,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.251, - "pct_cuda_time": 0.14434826086675084, + "cuda_time_us": 8.332, + "pct_cuda_time": 0.14379107977574668, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -12509,9 +12509,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 56.214, - "cuda_time_us": 35.803, - "pct_cuda_time": 0.626360536154682, + "cpu_time_us": 58.267, + "cuda_time_us": 35.081, + "pct_cuda_time": 0.6054170510817294, "trace": "" }, "children": [ @@ -12519,8 +12519,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.803, - "pct_cuda_time": 0.626360536154682, + "cuda_time_us": 35.081, + "pct_cuda_time": 0.6054170510817294, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -12534,18 +12534,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 549.931, - "cuda_time_us": 164.173, - "pct_cuda_time": 2.872147258668899, + "cpu_time_us": 612.346, + "cuda_time_us": 167.108, + "pct_cuda_time": 2.8838981948110267, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.743, + "cpu_time_us": 30.296, "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "pct_cuda_time": 0.08086953910575478, "trace": "" }, "children": [ @@ -12554,7 +12554,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "pct_cuda_time": 0.08086953910575478, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12564,18 +12564,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 342.361, - "cuda_time_us": 50.278, - "pct_cuda_time": 0.8795954259918191, + "cpu_time_us": 373.969, + "cuda_time_us": 54.125, + "pct_cuda_time": 0.9340725147458339, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 51.426, - "cuda_time_us": 16.954, - "pct_cuda_time": 0.2966040982589861, + "cpu_time_us": 53.35, + "cuda_time_us": 17.265, + "pct_cuda_time": 0.2979540317244679, "trace": "" }, "children": [ @@ -12583,8 +12583,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.954, - "pct_cuda_time": 0.2966040982589861, + "cuda_time_us": 17.265, + "pct_cuda_time": 0.2979540317244679, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12594,9 +12594,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.748, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.07006602651452395, + "cpu_time_us": 26.429, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06773643640420135, "trace": "" }, "children": [ @@ -12604,8 +12604,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.07006602651452395, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06773643640420135, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -12615,9 +12615,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 160.379, - "cuda_time_us": 14.698, - "pct_cuda_time": 0.2571361941848872, + "cpu_time_us": 181.742, + "cuda_time_us": 18.544, + "pct_cuda_time": 0.3200266182622955, "trace": "" }, "children": [ @@ -12625,8 +12625,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.923, - "pct_cuda_time": 0.05113682784068752, + "cuda_time_us": 5.767, + "pct_cuda_time": 0.09952510286446603, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12635,8 +12635,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.85, - "pct_cuda_time": 0.13733291089613311, + "cuda_time_us": 8.893, + "pct_cuda_time": 0.15347264431657648, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12645,8 +12645,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06866645544806656, + "cuda_time_us": 3.884, + "pct_cuda_time": 0.06702887108125301, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12656,9 +12656,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 58.417, - "cuda_time_us": 14.621, - "pct_cuda_time": 0.25578910703342195, + "cpu_time_us": 65.337, + "cuda_time_us": 14.391, + "pct_cuda_time": 0.2483554283548692, "trace": "" }, "children": [ @@ -12666,8 +12666,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.621, - "pct_cuda_time": 0.25578910703342195, + "cuda_time_us": 14.391, + "pct_cuda_time": 0.2483554283548692, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12679,9 +12679,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.511, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "cpu_time_us": 32.258, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "" }, "children": [ @@ -12689,8 +12689,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12700,18 +12700,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 133.469, - "cuda_time_us": 104.60300000000001, - "pct_cuda_time": 1.8299916533080522, + "cpu_time_us": 157.265, + "cuda_time_us": 103.61100000000002, + "pct_cuda_time": 1.7880866018536834, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 48.292, - "cuda_time_us": 61.641, - "pct_cuda_time": 1.078387001343763, + "cpu_time_us": 59.83, + "cuda_time_us": 61.591, + "pct_cuda_time": 1.0629184342856473, "trace": "" }, "children": [ @@ -12719,8 +12719,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 61.641, - "pct_cuda_time": 1.078387001343763, + "cuda_time_us": 61.591, + "pct_cuda_time": 1.0629184342856473, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12730,9 +12730,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 16.835, - "cuda_time_us": 8.211, - "pct_cuda_time": 0.14364847533352215, + "cpu_time_us": 20.391, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14032228392421942, "trace": "" }, "children": [ @@ -12740,8 +12740,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.211, - "pct_cuda_time": 0.14364847533352215, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14032228392421942, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -12751,9 +12751,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 53.169, - "cuda_time_us": 34.751, - "pct_cuda_time": 0.607956176630767, + "cpu_time_us": 58.507, + "cuda_time_us": 33.889, + "pct_cuda_time": 0.5848458836438165, "trace": "" }, "children": [ @@ -12761,8 +12761,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 34.751, - "pct_cuda_time": 0.607956176630767, + "cuda_time_us": 33.889, + "pct_cuda_time": 0.5848458836438165, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -12776,18 +12776,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 548.039, - "cuda_time_us": 163.683, - "pct_cuda_time": 2.8635748858868477, + "cpu_time_us": 583.132, + "cuda_time_us": 162.46200000000002, + "pct_cuda_time": 2.803718963337417, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 31.718, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.07918073308482781, + "cpu_time_us": 27.12, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "" }, "children": [ @@ -12795,8 +12795,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.07918073308482781, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12806,18 +12806,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 330.092, - "cuda_time_us": 47.655, - "pct_cuda_time": 0.833706989650347, + "cpu_time_us": 352.326, + "cuda_time_us": 48.046, + "pct_cuda_time": 0.829163012350639, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 57.526, - "cuda_time_us": 16.884, - "pct_cuda_time": 0.2953794735758358, + "cpu_time_us": 55.453, + "cuda_time_us": 16.664, + "pct_cuda_time": 0.28758215955149335, "trace": "" }, "children": [ @@ -12825,8 +12825,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.884, - "pct_cuda_time": 0.2953794735758358, + "cuda_time_us": 16.664, + "pct_cuda_time": 0.28758215955149335, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12836,9 +12836,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.937, + "cpu_time_us": 25.217, "cuda_time_us": 3.605, - "pct_cuda_time": 0.06306817118223691, + "pct_cuda_time": 0.06221397534704353, "trace": "" }, "children": [ @@ -12847,7 +12847,7 @@ "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.605, - "pct_cuda_time": 0.06306817118223691, + "pct_cuda_time": 0.06221397534704353, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -12857,9 +12857,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 154.03, - "cuda_time_us": 12.736, - "pct_cuda_time": 0.22281171378001927, + "cpu_time_us": 176.283, + "cuda_time_us": 13.737, + "pct_cuda_time": 0.2370688985693029, "trace": "" }, "children": [ @@ -12867,8 +12867,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.683, - "pct_cuda_time": 0.046938114641315296, + "cuda_time_us": 2.803, + "pct_cuda_time": 0.04837330732254175, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12877,8 +12877,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.128, - "pct_cuda_time": 0.10720714369063741, + "cuda_time_us": 7.049, + "pct_cuda_time": 0.12164946247470454, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12887,8 +12887,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06866645544806656, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06704612877205661, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12898,9 +12898,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 57.556, - "cuda_time_us": 14.43, - "pct_cuda_time": 0.2524476311122549, + "cpu_time_us": 61.001, + "cuda_time_us": 14.04, + "pct_cuda_time": 0.24229797888279922, "trace": "" }, "children": [ @@ -12908,8 +12908,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.43, - "pct_cuda_time": 0.2524476311122549, + "cuda_time_us": 14.04, + "pct_cuda_time": 0.24229797888279922, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12921,9 +12921,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.944, + "cpu_time_us": 31.578, "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "pct_cuda_time": 0.08086953910575478, "trace": "" }, "children": [ @@ -12932,7 +12932,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "pct_cuda_time": 0.08086953910575478, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12942,18 +12942,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 138.637, - "cuda_time_us": 106.816, - "pct_cuda_time": 1.8687072879339306, + "cpu_time_us": 149.413, + "cuda_time_us": 105.04400000000001, + "pct_cuda_time": 1.812816872775268, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 48.002, - "cuda_time_us": 64.706, - "pct_cuda_time": 1.1320080678274125, + "cpu_time_us": 52.619, + "cuda_time_us": 62.853, + "pct_cuda_time": 1.0846976400798134, "trace": "" }, "children": [ @@ -12961,8 +12961,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 64.706, - "pct_cuda_time": 1.1320080678274125, + "cuda_time_us": 62.853, + "pct_cuda_time": 1.0846976400798134, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12972,9 +12972,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 18.728, - "cuda_time_us": 8.331, - "pct_cuda_time": 0.14574783193320826, + "cpu_time_us": 22.574, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14170289918850887, "trace": "" }, "children": [ @@ -12982,8 +12982,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.331, - "pct_cuda_time": 0.14574783193320826, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14170289918850887, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -12993,9 +12993,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 56.475, - "cuda_time_us": 33.779, - "pct_cuda_time": 0.5909513881733096, + "cpu_time_us": 57.426, + "cuda_time_us": 33.98, + "pct_cuda_time": 0.5864163335069456, "trace": "" }, "children": [ @@ -13003,8 +13003,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.779, - "pct_cuda_time": 0.5909513881733096, + "cuda_time_us": 33.98, + "pct_cuda_time": 0.5864163335069456, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -13018,18 +13018,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 551.433, - "cuda_time_us": 163.773, - "pct_cuda_time": 2.8651494033366123, + "cpu_time_us": 562.09, + "cuda_time_us": 162.763, + "pct_cuda_time": 2.8089135282693056, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.98, - "cuda_time_us": 4.565, - "pct_cuda_time": 0.07986302397972582, + "cpu_time_us": 26.129, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "" }, "children": [ @@ -13037,8 +13037,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.565, - "pct_cuda_time": 0.07986302397972582, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13048,18 +13048,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 324.345, - "cuda_time_us": 48.43599999999999, - "pct_cuda_time": 0.8473703021866372, + "cpu_time_us": 340.609, + "cuda_time_us": 47.215, + "pct_cuda_time": 0.8148218712928322, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 48.593, - "cuda_time_us": 16.453, - "pct_cuda_time": 0.28783928445529655, + "cpu_time_us": 52.769, + "cuda_time_us": 17.085, + "pct_cuda_time": 0.2948476473798166, "trace": "" }, "children": [ @@ -13067,8 +13067,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.453, - "pct_cuda_time": 0.28783928445529655, + "cuda_time_us": 17.085, + "pct_cuda_time": 0.2948476473798166, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13078,9 +13078,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.729, - "cuda_time_us": 3.805, - "pct_cuda_time": 0.06656709884838044, + "cpu_time_us": 24.306, + "cuda_time_us": 3.684, + "pct_cuda_time": 0.06357733292052938, "trace": "" }, "children": [ @@ -13088,8 +13088,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.805, - "pct_cuda_time": 0.06656709884838044, + "cuda_time_us": 3.684, + "pct_cuda_time": 0.06357733292052938, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -13099,9 +13099,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 147.84, - "cuda_time_us": 14.097999999999999, - "pct_cuda_time": 0.2466394111864566, + "cpu_time_us": 168.722, + "cuda_time_us": 11.895, + "pct_cuda_time": 0.20528023210903823, "trace": "" }, "children": [ @@ -13109,8 +13109,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.603, - "pct_cuda_time": 0.0455385435748579, + "cuda_time_us": 2.523, + "pct_cuda_time": 0.04354115389752867, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13119,8 +13119,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.57, - "pct_cuda_time": 0.1324344121635322, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.11267546325682308, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13129,8 +13129,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06866645544806656, + "cuda_time_us": 2.843, + "pct_cuda_time": 0.04906361495468648, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13140,9 +13140,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 71.217, - "cuda_time_us": 14.08, - "pct_cuda_time": 0.2463245076965037, + "cpu_time_us": 58.998, + "cuda_time_us": 14.551, + "pct_cuda_time": 0.25111665888344814, "trace": "" }, "children": [ @@ -13150,8 +13150,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.08, - "pct_cuda_time": 0.2463245076965037, + "cuda_time_us": 14.551, + "pct_cuda_time": 0.25111665888344814, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13163,9 +13163,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.308, - "cuda_time_us": 4.806, - "pct_cuda_time": 0.08407923181742875, + "cpu_time_us": 31.056, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "" }, "children": [ @@ -13173,8 +13173,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.806, - "pct_cuda_time": 0.08407923181742875, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13184,18 +13184,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 148.803, - "cuda_time_us": 105.96600000000001, - "pct_cuda_time": 1.8538368453528205, + "cpu_time_us": 146.759, + "cuda_time_us": 106.29599999999999, + "pct_cuda_time": 1.834423501661398, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 58.828, - "cuda_time_us": 63.064, - "pct_cuda_time": 1.1032818716883743, + "cpu_time_us": 55.764, + "cuda_time_us": 62.202, + "pct_cuda_time": 1.073462883366658, "trace": "" }, "children": [ @@ -13203,8 +13203,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 63.064, - "pct_cuda_time": 1.1032818716883743, + "cuda_time_us": 62.202, + "pct_cuda_time": 1.073462883366658, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13214,9 +13214,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 17.156, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14154911873383605, + "cpu_time_us": 17.516, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.14101259155636414, "trace": "" }, "children": [ @@ -13224,8 +13224,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14154911873383605, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.14101259155636414, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -13235,9 +13235,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 57.626, - "cuda_time_us": 34.811, - "pct_cuda_time": 0.6090058549306101, + "cpu_time_us": 57.336, + "cuda_time_us": 35.923, + "pct_cuda_time": 0.6199480267383759, "trace": "" }, "children": [ @@ -13245,8 +13245,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 34.811, - "pct_cuda_time": 0.6090058549306101, + "cuda_time_us": 35.923, + "pct_cuda_time": 0.6199480267383759, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -13260,18 +13260,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 607.136, - "cuda_time_us": 162.099, - "pct_cuda_time": 2.835863378770991, + "cpu_time_us": 621.249, + "cuda_time_us": 163.25400000000002, + "pct_cuda_time": 2.8173870544538824, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.259, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07988051861805652, + "cpu_time_us": 30.025, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "" }, "children": [ @@ -13279,8 +13279,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07988051861805652, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13290,18 +13290,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 387.188, - "cuda_time_us": 46.68299999999999, - "pct_cuda_time": 0.8167022011928893, + "cpu_time_us": 393.218, + "cuda_time_us": 47.405, + "pct_cuda_time": 0.8181008325455198, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 47.741, - "cuda_time_us": 17.224, - "pct_cuda_time": 0.3013276506082798, + "cpu_time_us": 51.818, + "cuda_time_us": 17.465, + "pct_cuda_time": 0.3014055698851915, "trace": "" }, "children": [ @@ -13309,8 +13309,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.224, - "pct_cuda_time": 0.3013276506082798, + "cuda_time_us": 17.465, + "pct_cuda_time": 0.3014055698851915, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13320,9 +13320,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 23.255, - "cuda_time_us": 3.604, - "pct_cuda_time": 0.0630506765439062, + "cpu_time_us": 25.488, + "cuda_time_us": 3.645, + "pct_cuda_time": 0.06290428297918826, "trace": "" }, "children": [ @@ -13330,8 +13330,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.604, - "pct_cuda_time": 0.0630506765439062, + "cuda_time_us": 3.645, + "pct_cuda_time": 0.06290428297918826, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -13341,9 +13341,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 220.619, - "cuda_time_us": 11.936, - "pct_cuda_time": 0.20881600311544518, + "cpu_time_us": 222.182, + "cuda_time_us": 11.614, + "pct_cuda_time": 0.20043082099322154, "trace": "" }, "children": [ @@ -13352,7 +13352,7 @@ "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.683, - "pct_cuda_time": 0.046938114641315296, + "pct_cuda_time": 0.04630238442610757, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13361,8 +13361,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.728, - "pct_cuda_time": 0.10020928835835037, + "cuda_time_us": 6.208, + "pct_cuda_time": 0.10713574450886165, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13371,8 +13371,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.525, - "pct_cuda_time": 0.06166860011577951, + "cuda_time_us": 2.723, + "pct_cuda_time": 0.0469926920582523, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13382,9 +13382,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 61.843, - "cuda_time_us": 13.919, - "pct_cuda_time": 0.2435078709252582, + "cpu_time_us": 59.98, + "cuda_time_us": 14.681, + "pct_cuda_time": 0.25336015868791845, "trace": "" }, "children": [ @@ -13392,8 +13392,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.919, - "pct_cuda_time": 0.2435078709252582, + "cuda_time_us": 14.681, + "pct_cuda_time": 0.25336015868791845, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13405,9 +13405,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.624, + "cpu_time_us": 27.321, "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "pct_cuda_time": 0.08017923147361006, "trace": "" }, "children": [ @@ -13416,7 +13416,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "pct_cuda_time": 0.08017923147361006, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13426,18 +13426,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 147.21, - "cuda_time_us": 106.204, - "pct_cuda_time": 1.8580005692755313, + "cpu_time_us": 151.707, + "cuda_time_us": 106.59700000000001, + "pct_cuda_time": 1.8396180665932869, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 53.199, - "cuda_time_us": 65.326, - "pct_cuda_time": 1.1428547435924574, + "cpu_time_us": 53.349, + "cuda_time_us": 62.372, + "pct_cuda_time": 1.076396690803273, "trace": "" }, "children": [ @@ -13445,8 +13445,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 65.326, - "pct_cuda_time": 1.1428547435924574, + "cuda_time_us": 62.372, + "pct_cuda_time": 1.076396690803273, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13456,9 +13456,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 22.253, - "cuda_time_us": 8.131, - "pct_cuda_time": 0.14224890426706474, + "cpu_time_us": 26.189, + "cuda_time_us": 8.212, + "pct_cuda_time": 0.14172015687931247, "trace": "" }, "children": [ @@ -13466,8 +13466,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.131, - "pct_cuda_time": 0.14224890426706474, + "cuda_time_us": 8.212, + "pct_cuda_time": 0.14172015687931247, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -13477,9 +13477,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 53.189, - "cuda_time_us": 32.747, - "pct_cuda_time": 0.572896921416009, + "cpu_time_us": 55.553, + "cuda_time_us": 36.013, + "pct_cuda_time": 0.6215012189107014, "trace": "" }, "children": [ @@ -13487,8 +13487,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 32.747, - "pct_cuda_time": 0.572896921416009, + "cuda_time_us": 36.013, + "pct_cuda_time": 0.6215012189107014, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -13502,18 +13502,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 551.633, - "cuda_time_us": 160.287, - "pct_cuda_time": 2.804163094115731, + "cpu_time_us": 587.668, + "cuda_time_us": 167.377, + "pct_cuda_time": 2.8885405136372, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.813, + "cpu_time_us": 30.165, "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "pct_cuda_time": 0.07948892384146533, "trace": "" }, "children": [ @@ -13522,7 +13522,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "pct_cuda_time": 0.07948892384146533, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13532,18 +13532,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 334.499, - "cuda_time_us": 46.282, - "pct_cuda_time": 0.8096868512222716, + "cpu_time_us": 355.0, + "cuda_time_us": 50.828, + "pct_cuda_time": 0.8771739081663048, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 51.096, - "cuda_time_us": 16.884, - "pct_cuda_time": 0.2953794735758358, + "cpu_time_us": 52.709, + "cuda_time_us": 17.515, + "pct_cuda_time": 0.3022684544253724, "trace": "" }, "children": [ @@ -13551,8 +13551,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.884, - "pct_cuda_time": 0.2953794735758358, + "cuda_time_us": 17.515, + "pct_cuda_time": 0.3022684544253724, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13562,9 +13562,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.737, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06866645544806656, + "cpu_time_us": 26.48, + "cuda_time_us": 3.804, + "pct_cuda_time": 0.06564825581696355, "trace": "" }, "children": [ @@ -13572,8 +13572,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06866645544806656, + "cuda_time_us": 3.804, + "pct_cuda_time": 0.06564825581696355, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -13583,9 +13583,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 165.097, - "cuda_time_us": 12.255, - "pct_cuda_time": 0.2143967927429441, + "cpu_time_us": 175.663, + "cuda_time_us": 15.338999999999999, + "pct_cuda_time": 0.2647157192366992, "trace": "" }, "children": [ @@ -13593,8 +13593,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.643, - "pct_cuda_time": 0.04623832910808659, + "cuda_time_us": 2.602, + "pct_cuda_time": 0.0449045114710145, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13603,8 +13603,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.808, - "pct_cuda_time": 0.10160885942480777, + "cuda_time_us": 8.812, + "pct_cuda_time": 0.15207477136148337, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13613,8 +13613,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.804, - "pct_cuda_time": 0.06654960421004971, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06773643640420135, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13624,9 +13624,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 58.537, - "cuda_time_us": 13.218, - "pct_cuda_time": 0.23124412945542513, + "cpu_time_us": 61.762, + "cuda_time_us": 14.17, + "pct_cuda_time": 0.24454147868726958, "trace": "" }, "children": [ @@ -13634,8 +13634,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.218, - "pct_cuda_time": 0.23124412945542513, + "cuda_time_us": 14.17, + "pct_cuda_time": 0.24454147868726958, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13647,9 +13647,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.502, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "cpu_time_us": 29.695, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08225015437004424, "trace": "" }, "children": [ @@ -13657,8 +13657,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08225015437004424, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13668,18 +13668,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 143.574, - "cuda_time_us": 104.71300000000001, - "pct_cuda_time": 1.8319160635244314, + "cpu_time_us": 152.308, + "cuda_time_us": 107.17699999999999, + "pct_cuda_time": 1.8496275272593854, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 51.336, - "cuda_time_us": 61.811, - "pct_cuda_time": 1.081361089859985, + "cpu_time_us": 53.369, + "cuda_time_us": 63.103, + "pct_cuda_time": 1.0890120627807178, "trace": "" }, "children": [ @@ -13687,8 +13687,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 61.811, - "pct_cuda_time": 1.081361089859985, + "cuda_time_us": 63.103, + "pct_cuda_time": 1.0890120627807178, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13698,9 +13698,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 23.345, - "cuda_time_us": 8.211, - "pct_cuda_time": 0.14364847533352215, + "cpu_time_us": 26.5, + "cuda_time_us": 8.332, + "pct_cuda_time": 0.14379107977574668, "trace": "" }, "children": [ @@ -13708,8 +13708,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.211, - "pct_cuda_time": 0.14364847533352215, + "cuda_time_us": 8.332, + "pct_cuda_time": 0.14379107977574668, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -13719,9 +13719,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 52.969, - "cuda_time_us": 34.691, - "pct_cuda_time": 0.606906498330924, + "cpu_time_us": 53.34, + "cuda_time_us": 35.742, + "pct_cuda_time": 0.6168243847029209, "trace": "" }, "children": [ @@ -13729,8 +13729,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 34.691, - "pct_cuda_time": 0.606906498330924, + "cuda_time_us": 35.742, + "pct_cuda_time": 0.6168243847029209, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -13744,18 +13744,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 556.361, - "cuda_time_us": 164.135, - "pct_cuda_time": 2.871482462412332, + "cpu_time_us": 585.626, + "cuda_time_us": 165.216, + "pct_cuda_time": 2.851246643810581, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.902, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "cpu_time_us": 28.182, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08225015437004424, "trace": "" }, "children": [ @@ -13763,8 +13763,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08197987521774264, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08225015437004424, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13774,18 +13774,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 337.454, - "cuda_time_us": 50.028999999999996, - "pct_cuda_time": 0.8752392610474703, + "cpu_time_us": 358.146, + "cuda_time_us": 49.147, + "pct_cuda_time": 0.8481637299254225, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 49.394, - "cuda_time_us": 17.365, - "pct_cuda_time": 0.30379439461291097, + "cpu_time_us": 53.079, + "cuda_time_us": 16.704, + "pct_cuda_time": 0.288272467183638, "trace": "" }, "children": [ @@ -13793,8 +13793,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.365, - "pct_cuda_time": 0.30379439461291097, + "cuda_time_us": 16.704, + "pct_cuda_time": 0.288272467183638, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13804,9 +13804,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 28.082, - "cuda_time_us": 3.724, - "pct_cuda_time": 0.06515003314359231, + "cpu_time_us": 24.207, + "cuda_time_us": 3.725, + "pct_cuda_time": 0.06428489824347772, "trace": "" }, "children": [ @@ -13814,8 +13814,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.724, - "pct_cuda_time": 0.06515003314359231, + "cuda_time_us": 3.725, + "pct_cuda_time": 0.06428489824347772, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -13825,9 +13825,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 167.41, - "cuda_time_us": 14.780000000000001, - "pct_cuda_time": 0.25857075452800604, + "cpu_time_us": 187.3, + "cuda_time_us": 15.059000000000001, + "pct_cuda_time": 0.25988356581168615, "trace": "" }, "children": [ @@ -13835,8 +13835,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.964, - "pct_cuda_time": 0.051854108012246936, + "cuda_time_us": 2.963, + "pct_cuda_time": 0.05113453785112066, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13845,8 +13845,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.891, - "pct_cuda_time": 0.13805019106769253, + "cuda_time_us": 8.172, + "pct_cuda_time": 0.14102984924716777, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13855,8 +13855,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06866645544806656, + "cuda_time_us": 3.924, + "pct_cuda_time": 0.06771917871339773, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13866,9 +13866,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 59.229, - "cuda_time_us": 14.16, - "pct_cuda_time": 0.24772407876296113, + "cpu_time_us": 59.819, + "cuda_time_us": 13.659, + "pct_cuda_time": 0.23572279868662072, "trace": "" }, "children": [ @@ -13876,8 +13876,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.16, - "pct_cuda_time": 0.24772407876296113, + "cuda_time_us": 13.659, + "pct_cuda_time": 0.23572279868662072, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13889,9 +13889,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.2, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cpu_time_us": 27.541, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "" }, "children": [ @@ -13899,8 +13899,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13910,18 +13910,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 144.706, - "cuda_time_us": 104.77399999999999, - "pct_cuda_time": 1.8329832364626049, + "cpu_time_us": 149.564, + "cuda_time_us": 106.697, + "pct_cuda_time": 1.841343835673649, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 50.916, - "cuda_time_us": 61.361, - "pct_cuda_time": 1.073488502611162, + "cpu_time_us": 53.79, + "cuda_time_us": 62.983, + "pct_cuda_time": 1.0869411398842836, "trace": "" }, "children": [ @@ -13929,8 +13929,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 61.361, - "pct_cuda_time": 1.073488502611162, + "cuda_time_us": 62.983, + "pct_cuda_time": 1.0869411398842836, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13940,9 +13940,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.327, - "cuda_time_us": 8.131, - "pct_cuda_time": 0.14224890426706474, + "cpu_time_us": 22.473, + "cuda_time_us": 8.332, + "pct_cuda_time": 0.14379107977574668, "trace": "" }, "children": [ @@ -13950,8 +13950,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.131, - "pct_cuda_time": 0.14224890426706474, + "cuda_time_us": 8.332, + "pct_cuda_time": 0.14379107977574668, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -13961,9 +13961,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 53.48, - "cuda_time_us": 35.282, - "pct_cuda_time": 0.617245829584378, + "cpu_time_us": 53.32, + "cuda_time_us": 35.382, + "pct_cuda_time": 0.6106116160136184, "trace": "" }, "children": [ @@ -13971,8 +13971,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.282, - "pct_cuda_time": 0.617245829584378, + "cuda_time_us": 35.382, + "pct_cuda_time": 0.6106116160136184, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -13986,18 +13986,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 561.849, - "cuda_time_us": 162.751, - "pct_cuda_time": 2.847269882962619, + "cpu_time_us": 571.735, + "cuda_time_us": 164.063, + "pct_cuda_time": 2.831348526314009, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 30.075, - "cuda_time_us": 4.565, - "pct_cuda_time": 0.07986302397972582, + "cpu_time_us": 32.048, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "" }, "children": [ @@ -14005,8 +14005,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.565, - "pct_cuda_time": 0.07986302397972582, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14016,18 +14016,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 344.986, - "cuda_time_us": 49.628, - "pct_cuda_time": 0.8682239110768528, + "cpu_time_us": 344.235, + "cuda_time_us": 49.918, + "pct_cuda_time": 0.8614694095350123, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 51.186, - "cuda_time_us": 17.105, - "pct_cuda_time": 0.2992457886469244, + "cpu_time_us": 51.767, + "cuda_time_us": 16.364, + "pct_cuda_time": 0.28240485231040785, "trace": "" }, "children": [ @@ -14035,8 +14035,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.105, - "pct_cuda_time": 0.2992457886469244, + "cuda_time_us": 16.364, + "pct_cuda_time": 0.28240485231040785, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14046,9 +14046,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.408, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06936624098129525, + "cpu_time_us": 23.686, + "cuda_time_us": 3.884, + "pct_cuda_time": 0.06702887108125301, "trace": "" }, "children": [ @@ -14056,8 +14056,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06936624098129525, + "cuda_time_us": 3.884, + "pct_cuda_time": 0.06702887108125301, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -14067,9 +14067,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 175.091, - "cuda_time_us": 14.779, - "pct_cuda_time": 0.2585532598896753, + "cpu_time_us": 174.701, + "cuda_time_us": 15.46, + "pct_cuda_time": 0.26680389982393704, "trace": "" }, "children": [ @@ -14077,8 +14077,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.803, - "pct_cuda_time": 0.04903747124100141, + "cuda_time_us": 2.603, + "pct_cuda_time": 0.04492176916181812, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14087,8 +14087,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.931, - "pct_cuda_time": 0.13874997660092123, + "cuda_time_us": 9.012, + "pct_cuda_time": 0.15552630952220703, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14097,8 +14097,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.045, - "pct_cuda_time": 0.07076581204775266, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.0663558211399119, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14108,9 +14108,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 58.157, - "cuda_time_us": 13.779, - "pct_cuda_time": 0.24105862155895771, + "cpu_time_us": 59.549, + "cuda_time_us": 14.21, + "pct_cuda_time": 0.2452317863194143, "trace": "" }, "children": [ @@ -14118,8 +14118,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.779, - "pct_cuda_time": 0.24105862155895771, + "cuda_time_us": 14.21, + "pct_cuda_time": 0.2452317863194143, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14131,9 +14131,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.082, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cpu_time_us": 30.346, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "" }, "children": [ @@ -14141,8 +14141,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14152,18 +14152,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 142.052, - "cuda_time_us": 103.912, - "pct_cuda_time": 1.8179028582215266, + "cpu_time_us": 143.925, + "cuda_time_us": 104.81299999999999, + "pct_cuda_time": 1.808830346199632, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 51.898, - "cuda_time_us": 62.112, - "pct_cuda_time": 1.086626975997531, + "cpu_time_us": 53.911, + "cuda_time_us": 62.542, + "pct_cuda_time": 1.0793304982398881, "trace": "" }, "children": [ @@ -14171,8 +14171,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.112, - "pct_cuda_time": 1.086626975997531, + "cuda_time_us": 62.542, + "pct_cuda_time": 1.0793304982398881, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14182,9 +14182,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 21.182, - "cuda_time_us": 8.331, - "pct_cuda_time": 0.14574783193320826, + "cpu_time_us": 17.907, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.14101259155636414, "trace": "" }, "children": [ @@ -14192,8 +14192,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.331, - "pct_cuda_time": 0.14574783193320826, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.14101259155636414, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -14203,9 +14203,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 50.926, - "cuda_time_us": 33.469, - "pct_cuda_time": 0.5855280502907871, + "cpu_time_us": 54.312, + "cuda_time_us": 34.1, + "pct_cuda_time": 0.58848725640338, "trace": "" }, "children": [ @@ -14213,8 +14213,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.469, - "pct_cuda_time": 0.5855280502907871, + "cuda_time_us": 34.1, + "pct_cuda_time": 0.58848725640338, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -14228,18 +14228,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 518.975, - "cuda_time_us": 165.26600000000002, - "pct_cuda_time": 2.8912688983643737, + "cpu_time_us": 557.513, + "cuda_time_us": 163.03199999999998, + "pct_cuda_time": 2.8135558470954787, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 24.446, - "cuda_time_us": 4.525, - "pct_cuda_time": 0.07916323844649711, + "cpu_time_us": 27.972, + "cuda_time_us": 4.605, + "pct_cuda_time": 0.07947166615066172, "trace": "" }, "children": [ @@ -14247,8 +14247,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.525, - "pct_cuda_time": 0.07916323844649711, + "cuda_time_us": 4.605, + "pct_cuda_time": 0.07947166615066172, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14258,18 +14258,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 313.939, - "cuda_time_us": 47.696000000000005, - "pct_cuda_time": 0.8344242698219064, + "cpu_time_us": 330.754, + "cuda_time_us": 48.656, + "pct_cuda_time": 0.8396902037408461, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 52.408, - "cuda_time_us": 17.545, - "pct_cuda_time": 0.3069434295124402, + "cpu_time_us": 52.629, + "cuda_time_us": 17.054, + "pct_cuda_time": 0.2943126589649044, "trace": "" }, "children": [ @@ -14277,8 +14277,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.545, - "pct_cuda_time": 0.3069434295124402, + "cuda_time_us": 17.054, + "pct_cuda_time": 0.2943126589649044, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14288,9 +14288,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.757, - "cuda_time_us": 3.485, - "pct_cuda_time": 0.06096881458255081, + "cpu_time_us": 23.996, + "cuda_time_us": 3.524, + "pct_cuda_time": 0.060816102391950455, "trace": "" }, "children": [ @@ -14298,8 +14298,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.485, - "pct_cuda_time": 0.06096881458255081, + "cuda_time_us": 3.524, + "pct_cuda_time": 0.060816102391950455, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -14309,9 +14309,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 148.011, - "cuda_time_us": 12.536000000000001, - "pct_cuda_time": 0.21931278611387575, + "cpu_time_us": 163.064, + "cuda_time_us": 13.898000000000001, + "pct_cuda_time": 0.23984738678868547, "trace": "" }, "children": [ @@ -14319,8 +14319,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.563, - "pct_cuda_time": 0.04483875804162919, + "cuda_time_us": 2.683, + "pct_cuda_time": 0.04630238442610757, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14329,8 +14329,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.128, - "pct_cuda_time": 0.10720714369063741, + "cuda_time_us": 7.37, + "pct_cuda_time": 0.12718918122266595, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14340,7 +14340,7 @@ "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.845, - "pct_cuda_time": 0.06726688438160915, + "pct_cuda_time": 0.0663558211399119, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14350,9 +14350,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 56.324, - "cuda_time_us": 14.13, - "pct_cuda_time": 0.2471992396130396, + "cpu_time_us": 58.908, + "cuda_time_us": 14.18, + "pct_cuda_time": 0.24471405559530574, "trace": "" }, "children": [ @@ -14360,8 +14360,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.13, - "pct_cuda_time": 0.2471992396130396, + "cuda_time_us": 14.18, + "pct_cuda_time": 0.24471405559530574, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14373,9 +14373,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 25.368, + "cpu_time_us": 28.142, "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "pct_cuda_time": 0.08017923147361006, "trace": "" }, "children": [ @@ -14384,7 +14384,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "pct_cuda_time": 0.08017923147361006, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14394,18 +14394,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 137.355, - "cuda_time_us": 108.399, - "pct_cuda_time": 1.8964013004114564, + "cpu_time_us": 152.968, + "cuda_time_us": 105.125, + "pct_cuda_time": 1.8142147457303608, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 51.456, - "cuda_time_us": 64.706, - "pct_cuda_time": 1.1320080678274125, + "cpu_time_us": 59.278, + "cuda_time_us": 61.602, + "pct_cuda_time": 1.063108268884487, "trace": "" }, "children": [ @@ -14413,8 +14413,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 64.706, - "pct_cuda_time": 1.1320080678274125, + "cuda_time_us": 61.602, + "pct_cuda_time": 1.063108268884487, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14424,9 +14424,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 22.924, - "cuda_time_us": 8.171, - "pct_cuda_time": 0.14294868980029343, + "cpu_time_us": 21.452, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14170289918850887, "trace": "" }, "children": [ @@ -14434,8 +14434,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.171, - "pct_cuda_time": 0.14294868980029343, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14170289918850887, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -14445,9 +14445,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 48.342, - "cuda_time_us": 35.522, - "pct_cuda_time": 0.6214445427837503, + "cpu_time_us": 53.971, + "cuda_time_us": 35.312, + "pct_cuda_time": 0.6094035776573651, "trace": "" }, "children": [ @@ -14455,8 +14455,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.522, - "pct_cuda_time": 0.6214445427837503, + "cuda_time_us": 35.312, + "pct_cuda_time": 0.6094035776573651, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -14470,18 +14470,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 491.754, - "cuda_time_us": 163.332, - "pct_cuda_time": 2.8574342678327658, + "cpu_time_us": 557.994, + "cuda_time_us": 163.512, + "pct_cuda_time": 2.8218395386812154, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 23.144, - "cuda_time_us": 4.645, - "pct_cuda_time": 0.0812625950461832, + "cpu_time_us": 28.122, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "" }, "children": [ @@ -14489,8 +14489,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.645, - "pct_cuda_time": 0.0812625950461832, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14500,18 +14500,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 293.318, - "cuda_time_us": 48.40599999999999, - "pct_cuda_time": 0.8468454630367156, + "cpu_time_us": 341.009, + "cuda_time_us": 48.375, + "pct_cuda_time": 0.8348407926250294, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 48.943, - "cuda_time_us": 17.185, - "pct_cuda_time": 0.3006453597133818, + "cpu_time_us": 58.848, + "cuda_time_us": 16.754, + "pct_cuda_time": 0.289135351723819, "trace": "" }, "children": [ @@ -14519,8 +14519,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.185, - "pct_cuda_time": 0.3006453597133818, + "cuda_time_us": 16.754, + "pct_cuda_time": 0.289135351723819, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14530,9 +14530,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 21.552, - "cuda_time_us": 3.845, - "pct_cuda_time": 0.06726688438160915, + "cpu_time_us": 24.967, + "cuda_time_us": 3.804, + "pct_cuda_time": 0.06564825581696355, "trace": "" }, "children": [ @@ -14540,8 +14540,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.845, - "pct_cuda_time": 0.06726688438160915, + "cuda_time_us": 3.804, + "pct_cuda_time": 0.06564825581696355, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -14551,9 +14551,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 139.999, - "cuda_time_us": 13.255999999999998, - "pct_cuda_time": 0.23190892571199237, + "cpu_time_us": 168.472, + "cuda_time_us": 14.258000000000001, + "pct_cuda_time": 0.246060155477988, "trace": "" }, "children": [ @@ -14561,8 +14561,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.763, - "pct_cuda_time": 0.04833768570777271, + "cuda_time_us": 2.562, + "pct_cuda_time": 0.044214203838869766, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14571,8 +14571,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.649, - "pct_cuda_time": 0.11632185026094127, + "cuda_time_us": 7.851, + "pct_cuda_time": 0.13549013049920633, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14581,8 +14581,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.844, - "pct_cuda_time": 0.06724938974327843, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.0663558211399119, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14592,9 +14592,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 54.111, - "cuda_time_us": 14.12, - "pct_cuda_time": 0.2470242932297324, + "cpu_time_us": 59.419, + "cuda_time_us": 13.559, + "pct_cuda_time": 0.23399702960625887, "trace": "" }, "children": [ @@ -14602,8 +14602,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.12, - "pct_cuda_time": 0.2470242932297324, + "cuda_time_us": 13.559, + "pct_cuda_time": 0.23399702960625887, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14615,9 +14615,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 24.447, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "cpu_time_us": 28.212, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "" }, "children": [ @@ -14625,8 +14625,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14636,18 +14636,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 134.842, - "cuda_time_us": 105.675, - "pct_cuda_time": 1.8487459055985815, + "cpu_time_us": 143.544, + "cuda_time_us": 105.805, + "pct_cuda_time": 1.8259499754768216, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 51.106, - "cuda_time_us": 62.873, - "pct_cuda_time": 1.0999403957672071, + "cpu_time_us": 51.447, + "cuda_time_us": 62.683, + "pct_cuda_time": 1.0817638326431984, "trace": "" }, "children": [ @@ -14655,8 +14655,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.873, - "pct_cuda_time": 1.0999403957672071, + "cuda_time_us": 62.683, + "pct_cuda_time": 1.0817638326431984, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14666,9 +14666,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 17.055, - "cuda_time_us": 8.171, - "pct_cuda_time": 0.14294868980029343, + "cpu_time_us": 23.015, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14170289918850887, "trace": "" }, "children": [ @@ -14676,8 +14676,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.171, - "pct_cuda_time": 0.14294868980029343, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14170289918850887, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -14687,9 +14687,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 52.299, - "cuda_time_us": 34.631, - "pct_cuda_time": 0.605856820031081, + "cpu_time_us": 51.868, + "cuda_time_us": 34.911, + "pct_cuda_time": 0.6024832436451143, "trace": "" }, "children": [ @@ -14697,8 +14697,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 34.631, - "pct_cuda_time": 0.605856820031081, + "cuda_time_us": 34.911, + "pct_cuda_time": 0.6024832436451143, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -14712,18 +14712,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 564.262, - "cuda_time_us": 165.315, - "pct_cuda_time": 2.8921261356425787, + "cpu_time_us": 514.709, + "cuda_time_us": 165.125, + "pct_cuda_time": 2.849676193947452, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 23.966, - "cuda_time_us": 4.605, - "pct_cuda_time": 0.08056280951295451, + "cpu_time_us": 28.542, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "" }, "children": [ @@ -14731,8 +14731,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.605, - "pct_cuda_time": 0.08056280951295451, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14742,18 +14742,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 334.749, - "cuda_time_us": 48.846999999999994, - "pct_cuda_time": 0.8545605985405621, + "cpu_time_us": 312.177, + "cuda_time_us": 49.407000000000004, + "pct_cuda_time": 0.8526507295343634, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 47.872, - "cuda_time_us": 17.745, - "pct_cuda_time": 0.3104423571785837, + "cpu_time_us": 50.706, + "cuda_time_us": 16.684, + "pct_cuda_time": 0.2879273133675657, "trace": "" }, "children": [ @@ -14761,8 +14761,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.745, - "pct_cuda_time": 0.3104423571785837, + "cuda_time_us": 16.684, + "pct_cuda_time": 0.2879273133675657, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14772,9 +14772,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 23.836, - "cuda_time_us": 3.645, - "pct_cuda_time": 0.06376795671546563, + "cpu_time_us": 25.658, + "cuda_time_us": 3.685, + "pct_cuda_time": 0.06359459061133298, "trace": "" }, "children": [ @@ -14782,8 +14782,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.645, - "pct_cuda_time": 0.06376795671546563, + "cuda_time_us": 3.685, + "pct_cuda_time": 0.06359459061133298, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -14793,9 +14793,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 172.507, - "cuda_time_us": 13.818, - "pct_cuda_time": 0.24174091245385568, + "cpu_time_us": 147.891, + "cuda_time_us": 14.858, + "pct_cuda_time": 0.2564147699601589, "trace": "" }, "children": [ @@ -14803,8 +14803,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.803, - "pct_cuda_time": 0.04903747124100141, + "cuda_time_us": 2.883, + "pct_cuda_time": 0.04975392258683121, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14813,8 +14813,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.05, - "pct_cuda_time": 0.12333720023155902, + "cuda_time_us": 8.091, + "pct_cuda_time": 0.1396319762920747, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14823,8 +14823,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06936624098129525, + "cuda_time_us": 3.884, + "pct_cuda_time": 0.06702887108125301, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14834,9 +14834,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 59.278, - "cuda_time_us": 13.639, - "pct_cuda_time": 0.23860937219265724, + "cpu_time_us": 55.974, + "cuda_time_us": 14.18, + "pct_cuda_time": 0.24471405559530574, "trace": "" }, "children": [ @@ -14844,8 +14844,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.639, - "pct_cuda_time": 0.23860937219265724, + "cuda_time_us": 14.18, + "pct_cuda_time": 0.24471405559530574, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14857,9 +14857,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.79, - "cuda_time_us": 5.207, - "pct_cuda_time": 0.0910945817880465, + "cpu_time_us": 25.738, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "" }, "children": [ @@ -14867,8 +14867,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.207, - "pct_cuda_time": 0.0910945817880465, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14878,18 +14878,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 163.413, - "cuda_time_us": 106.656, - "pct_cuda_time": 1.8659081458010156, + "cpu_time_us": 131.376, + "cuda_time_us": 106.26599999999999, + "pct_cuda_time": 1.8339057709372892, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 54.391, - "cuda_time_us": 64.085, - "pct_cuda_time": 1.1211438974240369, + "cpu_time_us": 52.779, + "cuda_time_us": 63.354, + "pct_cuda_time": 1.0933437431724262, "trace": "" }, "children": [ @@ -14897,8 +14897,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 64.085, - "pct_cuda_time": 1.1211438974240369, + "cuda_time_us": 63.354, + "pct_cuda_time": 1.0933437431724262, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14908,9 +14908,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 33.42, - "cuda_time_us": 8.451, - "pct_cuda_time": 0.14784718853289439, + "cpu_time_us": 17.536, + "cuda_time_us": 8.692, + "pct_cuda_time": 0.15000384846504922, "trace": "" }, "children": [ @@ -14918,8 +14918,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.451, - "pct_cuda_time": 0.14784718853289439, + "cuda_time_us": 8.692, + "pct_cuda_time": 0.15000384846504922, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -14929,9 +14929,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 58.547, - "cuda_time_us": 34.12, - "pct_cuda_time": 0.5969170598440843, + "cpu_time_us": 46.019, + "cuda_time_us": 34.22, + "pct_cuda_time": 0.590558179299814, "trace": "" }, "children": [ @@ -14939,8 +14939,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 34.12, - "pct_cuda_time": 0.5969170598440843, + "cuda_time_us": 34.22, + "pct_cuda_time": 0.590558179299814, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -14954,18 +14954,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 560.597, - "cuda_time_us": 162.92100000000002, - "pct_cuda_time": 2.8502439714788412, + "cpu_time_us": 544.103, + "cuda_time_us": 165.72899999999998, + "pct_cuda_time": 2.860099839192837, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 30.255, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "cpu_time_us": 24.256, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "" }, "children": [ @@ -14973,8 +14973,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14984,18 +14984,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 342.251, - "cuda_time_us": 47.745000000000005, - "pct_cuda_time": 0.8352815071001116, + "cpu_time_us": 329.723, + "cuda_time_us": 49.379999999999995, + "pct_cuda_time": 0.8521847718826655, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 55.924, - "cuda_time_us": 16.183, - "pct_cuda_time": 0.28311573210600277, + "cpu_time_us": 47.791, + "cuda_time_us": 17.546, + "pct_cuda_time": 0.30280344284028454, "trace": "" }, "children": [ @@ -15003,8 +15003,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.183, - "pct_cuda_time": 0.28311573210600277, + "cuda_time_us": 17.546, + "pct_cuda_time": 0.30280344284028454, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15014,9 +15014,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.989, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.07006602651452395, + "cpu_time_us": 25.849, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06773643640420135, "trace": "" }, "children": [ @@ -15024,8 +15024,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.07006602651452395, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06773643640420135, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -15035,9 +15035,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 169.593, - "cuda_time_us": 13.577, - "pct_cuda_time": 0.23752470461615274, + "cpu_time_us": 165.127, + "cuda_time_us": 13.578, + "pct_cuda_time": 0.23432492573152763, "trace": "" }, "children": [ @@ -15045,8 +15045,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.843, - "pct_cuda_time": 0.04973725677423011, + "cuda_time_us": 2.723, + "pct_cuda_time": 0.0469926920582523, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15055,8 +15055,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.889, - "pct_cuda_time": 0.1205205634603135, + "cuda_time_us": 7.09, + "pct_cuda_time": 0.12235702779765287, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15065,8 +15065,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.845, - "pct_cuda_time": 0.06726688438160915, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06497520587562244, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15076,9 +15076,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 58.667, - "cuda_time_us": 13.98, - "pct_cuda_time": 0.24457504386343193, + "cpu_time_us": 58.257, + "cuda_time_us": 14.331, + "pct_cuda_time": 0.24731996690665212, "trace": "" }, "children": [ @@ -15086,8 +15086,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.98, - "pct_cuda_time": 0.24457504386343193, + "cuda_time_us": 14.331, + "pct_cuda_time": 0.24731996690665212, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15099,9 +15099,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.803, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cpu_time_us": 28.152, + "cuda_time_us": 5.127, + "pct_cuda_time": 0.08848018075015039, "trace": "" }, "children": [ @@ -15109,8 +15109,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cuda_time_us": 5.127, + "pct_cuda_time": 0.08848018075015039, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15120,18 +15120,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 142.082, - "cuda_time_us": 105.924, - "pct_cuda_time": 1.8531020705429304, + "cpu_time_us": 146.058, + "cuda_time_us": 106.536, + "pct_cuda_time": 1.8385653474542665, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 54.121, - "cuda_time_us": 63.674, - "pct_cuda_time": 1.1139536010701119, + "cpu_time_us": 54.852, + "cuda_time_us": 62.893, + "pct_cuda_time": 1.0853879477119581, "trace": "" }, "children": [ @@ -15139,8 +15139,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 63.674, - "pct_cuda_time": 1.1139536010701119, + "cuda_time_us": 62.893, + "pct_cuda_time": 1.0853879477119581, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15150,9 +15150,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 20.771, - "cuda_time_us": 8.371, - "pct_cuda_time": 0.14644761746643697, + "cpu_time_us": 23.895, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.14308351445279832, "trace": "" }, "children": [ @@ -15160,8 +15160,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.371, - "pct_cuda_time": 0.14644761746643697, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.14308351445279832, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -15171,9 +15171,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 51.227, - "cuda_time_us": 33.879, - "pct_cuda_time": 0.5927008520063812, + "cpu_time_us": 49.855, + "cuda_time_us": 35.352, + "pct_cuda_time": 0.6100938852895098, "trace": "" }, "children": [ @@ -15181,8 +15181,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.879, - "pct_cuda_time": 0.5927008520063812, + "cuda_time_us": 35.352, + "pct_cuda_time": 0.6100938852895098, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -15196,18 +15196,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 597.322, - "cuda_time_us": 162.149, - "pct_cuda_time": 2.836738110687527, + "cpu_time_us": 633.147, + "cuda_time_us": 165.76600000000002, + "pct_cuda_time": 2.8607383737525716, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.131, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07988051861805652, + "cpu_time_us": 26.029, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "" }, "children": [ @@ -15215,8 +15215,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07988051861805652, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15226,18 +15226,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 346.618, - "cuda_time_us": 49.016, - "pct_cuda_time": 0.8575171924184536, + "cpu_time_us": 344.455, + "cuda_time_us": 49.097, + "pct_cuda_time": 0.8473008453852418, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 51.387, - "cuda_time_us": 17.535, - "pct_cuda_time": 0.30676848312913296, + "cpu_time_us": 49.734, + "cuda_time_us": 16.664, + "pct_cuda_time": 0.28758215955149335, "trace": "" }, "children": [ @@ -15245,8 +15245,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.535, - "pct_cuda_time": 0.30676848312913296, + "cuda_time_us": 16.664, + "pct_cuda_time": 0.28758215955149335, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15256,9 +15256,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 23.796, - "cuda_time_us": 3.564, - "pct_cuda_time": 0.062350891010677506, + "cpu_time_us": 23.475, + "cuda_time_us": 3.604, + "pct_cuda_time": 0.06219671765623992, "trace": "" }, "children": [ @@ -15266,8 +15266,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.564, - "pct_cuda_time": 0.062350891010677506, + "cuda_time_us": 3.604, + "pct_cuda_time": 0.06219671765623992, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -15277,9 +15277,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 183.123, - "cuda_time_us": 13.737, - "pct_cuda_time": 0.24032384674906757, + "cpu_time_us": 181.241, + "cuda_time_us": 14.058, + "pct_cuda_time": 0.24260861731726432, "trace": "" }, "children": [ @@ -15287,8 +15287,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.683, - "pct_cuda_time": 0.046938114641315296, + "cuda_time_us": 2.763, + "pct_cuda_time": 0.04768299969039703, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15297,8 +15297,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.089, - "pct_cuda_time": 0.12401949112645701, + "cuda_time_us": 7.41, + "pct_cuda_time": 0.1278794888548107, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15307,8 +15307,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06936624098129525, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06704612877205661, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15318,9 +15318,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 58.608, - "cuda_time_us": 14.18, - "pct_cuda_time": 0.24807397152957544, + "cpu_time_us": 58.878, + "cuda_time_us": 14.771, + "pct_cuda_time": 0.25491335086024414, "trace": "" }, "children": [ @@ -15328,8 +15328,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.18, - "pct_cuda_time": 0.24807397152957544, + "cuda_time_us": 14.771, + "pct_cuda_time": 0.25491335086024414, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15341,9 +15341,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 73.009, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cpu_time_us": 87.441, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.08363076963433369, "trace": "" }, "children": [ @@ -15351,8 +15351,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.08363076963433369, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15362,18 +15362,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 134.641, - "cuda_time_us": 103.921, - "pct_cuda_time": 1.818060309966503, + "cpu_time_us": 157.325, + "cuda_time_us": 107.137, + "pct_cuda_time": 1.848937219627241, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 54.031, - "cuda_time_us": 62.151, - "pct_cuda_time": 1.0873092668924291, + "cpu_time_us": 64.537, + "cuda_time_us": 63.914, + "pct_cuda_time": 1.1030080500224522, "trace": "" }, "children": [ @@ -15381,8 +15381,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.151, - "pct_cuda_time": 1.0873092668924291, + "cuda_time_us": 63.914, + "pct_cuda_time": 1.1030080500224522, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15392,9 +15392,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 17.857, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14154911873383605, + "cpu_time_us": 22.544, + "cuda_time_us": 8.252, + "pct_cuda_time": 0.14241046451145722, "trace": "" }, "children": [ @@ -15402,8 +15402,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14154911873383605, + "cuda_time_us": 8.252, + "pct_cuda_time": 0.14241046451145722, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -15413,9 +15413,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 46.77, - "cuda_time_us": 33.679, - "pct_cuda_time": 0.5892019243402378, + "cpu_time_us": 53.54, + "cuda_time_us": 34.971, + "pct_cuda_time": 0.6035187050933313, "trace": "" }, "children": [ @@ -15423,8 +15423,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.679, - "pct_cuda_time": 0.5892019243402378, + "cuda_time_us": 34.971, + "pct_cuda_time": 0.6035187050933313, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -15438,18 +15438,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 522.741, - "cuda_time_us": 164.044, - "pct_cuda_time": 2.869890450324237, + "cpu_time_us": 602.471, + "cuda_time_us": 169.733, + "pct_cuda_time": 2.9291996331705246, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 25.798, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "cpu_time_us": 31.737, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "" }, "children": [ @@ -15457,8 +15457,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15468,18 +15468,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 306.628, - "cuda_time_us": 47.935, - "pct_cuda_time": 0.8386054883829479, + "cpu_time_us": 374.55, + "cuda_time_us": 53.21300000000001, + "pct_cuda_time": 0.9183335007329343, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 49.153, - "cuda_time_us": 17.155, - "pct_cuda_time": 0.30012052056346034, + "cpu_time_us": 57.586, + "cuda_time_us": 17.695, + "pct_cuda_time": 0.3053748387700237, "trace": "" }, "children": [ @@ -15487,8 +15487,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.155, - "pct_cuda_time": 0.30012052056346034, + "cuda_time_us": 17.695, + "pct_cuda_time": 0.3053748387700237, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15498,9 +15498,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 28.433, - "cuda_time_us": 3.804, - "pct_cuda_time": 0.06654960421004971, + "cpu_time_us": 25.979, + "cuda_time_us": 3.724, + "pct_cuda_time": 0.0642676405526741, "trace": "" }, "children": [ @@ -15508,8 +15508,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.804, - "pct_cuda_time": 0.06654960421004971, + "cuda_time_us": 3.724, + "pct_cuda_time": 0.0642676405526741, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -15519,9 +15519,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 144.716, - "cuda_time_us": 13.136, - "pct_cuda_time": 0.2298095691123063, + "cpu_time_us": 191.387, + "cuda_time_us": 17.624000000000002, + "pct_cuda_time": 0.30414954272296685, "trace": "" }, "children": [ @@ -15529,8 +15529,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.603, - "pct_cuda_time": 0.0455385435748579, + "cuda_time_us": 6.008, + "pct_cuda_time": 0.10368420634813802, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15539,8 +15539,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.608, - "pct_cuda_time": 0.11560457008938185, + "cuda_time_us": 7.731, + "pct_cuda_time": 0.13341920760277212, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15549,8 +15549,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06866645544806656, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06704612877205661, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15560,9 +15560,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 56.103, - "cuda_time_us": 13.84, - "pct_cuda_time": 0.24212579449713148, + "cpu_time_us": 63.515, + "cuda_time_us": 14.17, + "pct_cuda_time": 0.24454147868726958, "trace": "" }, "children": [ @@ -15570,8 +15570,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.84, - "pct_cuda_time": 0.24212579449713148, + "cuda_time_us": 14.17, + "pct_cuda_time": 0.24454147868726958, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15583,9 +15583,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.811, - "cuda_time_us": 5.127, - "pct_cuda_time": 0.08969501072158909, + "cpu_time_us": 31.006, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "" }, "children": [ @@ -15593,8 +15593,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.127, - "pct_cuda_time": 0.08969501072158909, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15604,18 +15604,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 144.606, - "cuda_time_us": 106.376, - "pct_cuda_time": 1.8610096470684148, + "cpu_time_us": 146.519, + "cuda_time_us": 107.268, + "pct_cuda_time": 1.8511979771225147, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 51.687, - "cuda_time_us": 62.132, - "pct_cuda_time": 1.0869768687641455, + "cpu_time_us": 53.43, + "cuda_time_us": 63.484, + "pct_cuda_time": 1.0955872429768965, "trace": "" }, "children": [ @@ -15623,8 +15623,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.132, - "pct_cuda_time": 1.0869768687641455, + "cuda_time_us": 63.484, + "pct_cuda_time": 1.0955872429768965, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15634,9 +15634,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 22.244, - "cuda_time_us": 8.291, - "pct_cuda_time": 0.14504804639997956, + "cpu_time_us": 23.225, + "cuda_time_us": 8.172, + "pct_cuda_time": 0.14102984924716777, "trace": "" }, "children": [ @@ -15644,8 +15644,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.291, - "pct_cuda_time": 0.14504804639997956, + "cuda_time_us": 8.172, + "pct_cuda_time": 0.14102984924716777, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -15655,9 +15655,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 53.49, - "cuda_time_us": 35.953, - "pct_cuda_time": 0.6289847319042896, + "cpu_time_us": 52.919, + "cuda_time_us": 35.612, + "pct_cuda_time": 0.6145808848984506, "trace": "" }, "children": [ @@ -15665,8 +15665,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.953, - "pct_cuda_time": 0.6289847319042896, + "cuda_time_us": 35.612, + "pct_cuda_time": 0.6145808848984506, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -15680,18 +15680,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 493.928, - "cuda_time_us": 161.761, - "pct_cuda_time": 2.8299501910152083, + "cpu_time_us": 570.653, + "cuda_time_us": 166.458, + "pct_cuda_time": 2.8726806957886746, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.239, + "cpu_time_us": 28.433, "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "pct_cuda_time": 0.07948892384146533, "trace": "" }, "children": [ @@ -15700,7 +15700,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "pct_cuda_time": 0.07948892384146533, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15710,18 +15710,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 295.01, - "cuda_time_us": 47.916, - "pct_cuda_time": 0.8382730902546641, + "cpu_time_us": 351.806, + "cuda_time_us": 49.778, + "pct_cuda_time": 0.8590533328225057, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 47.1, - "cuda_time_us": 17.695, - "pct_cuda_time": 0.30956762526204784, + "cpu_time_us": 59.64, + "cuda_time_us": 17.555, + "pct_cuda_time": 0.30295876205751715, "trace": "" }, "children": [ @@ -15729,8 +15729,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.695, - "pct_cuda_time": 0.30956762526204784, + "cuda_time_us": 17.555, + "pct_cuda_time": 0.30295876205751715, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15740,9 +15740,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 23.455, - "cuda_time_us": 3.605, - "pct_cuda_time": 0.06306817118223691, + "cpu_time_us": 27.381, + "cuda_time_us": 3.644, + "pct_cuda_time": 0.06288702528838465, "trace": "" }, "children": [ @@ -15750,8 +15750,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.605, - "pct_cuda_time": 0.06306817118223691, + "cuda_time_us": 3.644, + "pct_cuda_time": 0.06288702528838465, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -15761,9 +15761,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 139.388, - "cuda_time_us": 12.616, - "pct_cuda_time": 0.22071235718033314, + "cpu_time_us": 170.115, + "cuda_time_us": 14.058, + "pct_cuda_time": 0.24260861731726432, "trace": "" }, "children": [ @@ -15771,8 +15771,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.643, - "pct_cuda_time": 0.04623832910808659, + "cuda_time_us": 2.683, + "pct_cuda_time": 0.04630238442610757, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15781,8 +15781,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.008, - "pct_cuda_time": 0.1051077870909513, + "cuda_time_us": 7.49, + "pct_cuda_time": 0.12926010411910016, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15791,8 +15791,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06936624098129525, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06704612877205661, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15802,9 +15802,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 54.622, - "cuda_time_us": 14.0, - "pct_cuda_time": 0.24492493663004627, + "cpu_time_us": 58.477, + "cuda_time_us": 14.521, + "pct_cuda_time": 0.2505989281593396, "trace": "" }, "children": [ @@ -15812,8 +15812,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.0, - "pct_cuda_time": 0.24492493663004627, + "cuda_time_us": 14.521, + "pct_cuda_time": 0.2505989281593396, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15825,9 +15825,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 25.438, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cpu_time_us": 29.534, + "cuda_time_us": 5.007, + "pct_cuda_time": 0.08640925785371621, "trace": "" }, "children": [ @@ -15835,8 +15835,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cuda_time_us": 5.007, + "pct_cuda_time": 0.08640925785371621, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15846,18 +15846,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 131.596, - "cuda_time_us": 104.59299999999999, - "pct_cuda_time": 1.8298167069247449, + "cpu_time_us": 144.206, + "cuda_time_us": 107.067, + "pct_cuda_time": 1.8477291812709875, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 51.537, - "cuda_time_us": 63.053, - "pct_cuda_time": 1.1030894306667363, + "cpu_time_us": 52.478, + "cuda_time_us": 64.535, + "pct_cuda_time": 1.1137250760114992, "trace": "" }, "children": [ @@ -15865,8 +15865,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 63.053, - "pct_cuda_time": 1.1030894306667363, + "cuda_time_us": 64.535, + "pct_cuda_time": 1.1137250760114992, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15876,9 +15876,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 17.937, - "cuda_time_us": 8.171, - "pct_cuda_time": 0.14294868980029343, + "cpu_time_us": 23.124, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.14308351445279832, "trace": "" }, "children": [ @@ -15886,8 +15886,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.171, - "pct_cuda_time": 0.14294868980029343, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.14308351445279832, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -15897,9 +15897,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 47.451, - "cuda_time_us": 33.369, - "pct_cuda_time": 0.5837785864577154, + "cpu_time_us": 51.617, + "cuda_time_us": 34.241, + "pct_cuda_time": 0.59092059080669, "trace": "" }, "children": [ @@ -15907,8 +15907,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.369, - "pct_cuda_time": 0.5837785864577154, + "cuda_time_us": 34.241, + "pct_cuda_time": 0.59092059080669, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -15922,18 +15922,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 506.266, - "cuda_time_us": 161.741, - "pct_cuda_time": 2.8296002982485944, + "cpu_time_us": 580.658, + "cuda_time_us": 166.118, + "pct_cuda_time": 2.8668130809154446, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 24.086, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "cpu_time_us": 27.681, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "" }, "children": [ @@ -15941,8 +15941,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15952,18 +15952,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 312.476, - "cuda_time_us": 48.846999999999994, - "pct_cuda_time": 0.8545605985405621, + "cpu_time_us": 350.534, + "cuda_time_us": 49.418000000000006, + "pct_cuda_time": 0.8528405641332032, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 47.321, - "cuda_time_us": 17.104, - "pct_cuda_time": 0.2992282940085937, + "cpu_time_us": 52.258, + "cuda_time_us": 17.385, + "pct_cuda_time": 0.30002495462090206, "trace": "" }, "children": [ @@ -15971,8 +15971,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.104, - "pct_cuda_time": 0.2992282940085937, + "cuda_time_us": 17.385, + "pct_cuda_time": 0.30002495462090206, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15982,9 +15982,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 23.094, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.07006602651452395, + "cpu_time_us": 30.846, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06704612877205661, "trace": "" }, "children": [ @@ -15992,8 +15992,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.07006602651452395, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06704612877205661, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -16003,9 +16003,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 158.337, - "cuda_time_us": 12.737, - "pct_cuda_time": 0.22282920841834997, + "cpu_time_us": 174.141, + "cuda_time_us": 13.978, + "pct_cuda_time": 0.2412280020529749, "trace": "" }, "children": [ @@ -16013,8 +16013,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.844, - "pct_cuda_time": 0.049754751412560826, + "cuda_time_us": 2.643, + "pct_cuda_time": 0.04561207679396284, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16023,8 +16023,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.048, - "pct_cuda_time": 0.10580757262418002, + "cuda_time_us": 7.451, + "pct_cuda_time": 0.12858705417775904, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16033,8 +16033,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.845, - "pct_cuda_time": 0.06726688438160915, + "cuda_time_us": 3.884, + "pct_cuda_time": 0.06702887108125301, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16044,9 +16044,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 56.284, - "cuda_time_us": 15.001, - "pct_cuda_time": 0.2624370695990946, + "cpu_time_us": 59.078, + "cuda_time_us": 14.17, + "pct_cuda_time": 0.24454147868726958, "trace": "" }, "children": [ @@ -16054,8 +16054,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.001, - "pct_cuda_time": 0.2624370695990946, + "cuda_time_us": 14.17, + "pct_cuda_time": 0.24454147868726958, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16067,9 +16067,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 25.027, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.08267966075097134, + "cpu_time_us": 30.375, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08294046200218896, "trace": "" }, "children": [ @@ -16077,8 +16077,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.08267966075097134, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08294046200218896, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16088,18 +16088,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 128.933, - "cuda_time_us": 103.56200000000001, - "pct_cuda_time": 1.8117797348057754, + "cpu_time_us": 153.259, + "cuda_time_us": 107.168, + "pct_cuda_time": 1.8494722080421533, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 48.092, - "cuda_time_us": 62.042, - "pct_cuda_time": 1.085402351314381, + "cpu_time_us": 60.831, + "cuda_time_us": 65.397, + "pct_cuda_time": 1.128601205484218, "trace": "" }, "children": [ @@ -16107,8 +16107,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.042, - "pct_cuda_time": 1.085402351314381, + "cuda_time_us": 65.397, + "pct_cuda_time": 1.128601205484218, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16118,9 +16118,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 18.798, - "cuda_time_us": 8.251, - "pct_cuda_time": 0.14434826086675084, + "cpu_time_us": 21.352, + "cuda_time_us": 8.332, + "pct_cuda_time": 0.14379107977574668, "trace": "" }, "children": [ @@ -16128,8 +16128,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.251, - "pct_cuda_time": 0.14434826086675084, + "cuda_time_us": 8.332, + "pct_cuda_time": 0.14379107977574668, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -16139,9 +16139,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 47.511, - "cuda_time_us": 33.269, - "pct_cuda_time": 0.5820291226246436, + "cpu_time_us": 53.34, + "cuda_time_us": 33.439, + "pct_cuda_time": 0.5770799227821883, "trace": "" }, "children": [ @@ -16149,8 +16149,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.269, - "pct_cuda_time": 0.5820291226246436, + "cuda_time_us": 33.439, + "pct_cuda_time": 0.5770799227821883, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -16164,18 +16164,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 474.929, - "cuda_time_us": 164.305, - "pct_cuda_time": 2.8744565509285542, + "cpu_time_us": 574.578, + "cuda_time_us": 165.685, + "pct_cuda_time": 2.859340500797478, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 23.335, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cpu_time_us": 29.394, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08225015437004424, "trace": "" }, "children": [ @@ -16183,8 +16183,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08225015437004424, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16194,18 +16194,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 290.173, - "cuda_time_us": 49.408, - "pct_cuda_time": 0.8643750906440949, + "cpu_time_us": 347.77, + "cuda_time_us": 49.237, + "pct_cuda_time": 0.8497169220977483, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 46.058, - "cuda_time_us": 17.305, - "pct_cuda_time": 0.30274471631306793, + "cpu_time_us": 50.826, + "cuda_time_us": 16.854, + "pct_cuda_time": 0.29086112080418075, "trace": "" }, "children": [ @@ -16213,8 +16213,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.305, - "pct_cuda_time": 0.30274471631306793, + "cuda_time_us": 16.854, + "pct_cuda_time": 0.29086112080418075, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16224,9 +16224,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 21.532, + "cpu_time_us": 24.065, "cuda_time_us": 3.765, - "pct_cuda_time": 0.06586731331515173, + "pct_cuda_time": 0.06497520587562244, "trace": "" }, "children": [ @@ -16235,7 +16235,7 @@ "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.765, - "pct_cuda_time": 0.06586731331515173, + "pct_cuda_time": 0.06497520587562244, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -16245,9 +16245,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 142.583, - "cuda_time_us": 13.657, - "pct_cuda_time": 0.23892427568261015, + "cpu_time_us": 164.255, + "cuda_time_us": 14.658, + "pct_cuda_time": 0.25296323179943525, "trace": "" }, "children": [ @@ -16255,8 +16255,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.923, - "pct_cuda_time": 0.05113682784068752, + "cuda_time_us": 3.003, + "pct_cuda_time": 0.05182484548326539, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16265,8 +16265,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.809, - "pct_cuda_time": 0.1191209923938561, + "cuda_time_us": 7.81, + "pct_cuda_time": 0.13478256517625797, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16275,8 +16275,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06866645544806656, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.0663558211399119, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16286,9 +16286,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 54.211, - "cuda_time_us": 14.681, - "pct_cuda_time": 0.25683878533326493, + "cpu_time_us": 74.171, + "cuda_time_us": 13.96, + "pct_cuda_time": 0.24091736361850977, "trace": "" }, "children": [ @@ -16296,8 +16296,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.681, - "pct_cuda_time": 0.25683878533326493, + "cuda_time_us": 13.96, + "pct_cuda_time": 0.24091736361850977, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16309,9 +16309,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 25.668, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cpu_time_us": 30.265, + "cuda_time_us": 4.926, + "pct_cuda_time": 0.08501138489862314, "trace": "" }, "children": [ @@ -16319,8 +16319,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cuda_time_us": 4.926, + "pct_cuda_time": 0.08501138489862314, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16330,18 +16330,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 120.44, - "cuda_time_us": 105.605, - "pct_cuda_time": 1.8475212809154313, + "cpu_time_us": 150.045, + "cuda_time_us": 106.756, + "pct_cuda_time": 1.8423620394310625, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 45.438, - "cuda_time_us": 61.601, - "pct_cuda_time": 1.0776872158105344, + "cpu_time_us": 57.366, + "cuda_time_us": 64.766, + "pct_cuda_time": 1.1177116025871352, "trace": "" }, "children": [ @@ -16349,8 +16349,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 61.601, - "pct_cuda_time": 1.0776872158105344, + "cuda_time_us": 64.766, + "pct_cuda_time": 1.1177116025871352, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16360,9 +16360,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 16.776, - "cuda_time_us": 8.292, - "pct_cuda_time": 0.1450655410383103, + "cpu_time_us": 21.932, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14170289918850887, "trace": "" }, "children": [ @@ -16370,8 +16370,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.292, - "pct_cuda_time": 0.1450655410383103, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14170289918850887, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -16381,9 +16381,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 44.667, - "cuda_time_us": 35.712, - "pct_cuda_time": 0.6247685240665867, + "cpu_time_us": 53.65, + "cuda_time_us": 33.779, + "pct_cuda_time": 0.5829475376554185, "trace": "" }, "children": [ @@ -16391,8 +16391,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.712, - "pct_cuda_time": 0.6247685240665867, + "cuda_time_us": 33.779, + "pct_cuda_time": 0.5829475376554185, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -16406,18 +16406,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 471.865, - "cuda_time_us": 161.663, - "pct_cuda_time": 2.8282357164587983, + "cpu_time_us": 563.141, + "cuda_time_us": 163.474, + "pct_cuda_time": 2.821183746430678, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 22.854, - "cuda_time_us": 4.406, - "pct_cuda_time": 0.07708137648514171, + "cpu_time_us": 28.512, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "" }, "children": [ @@ -16425,8 +16425,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.406, - "pct_cuda_time": 0.07708137648514171, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16436,18 +16436,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 285.416, - "cuda_time_us": 47.747, - "pct_cuda_time": 0.8353164963767729, + "cpu_time_us": 337.555, + "cuda_time_us": 48.717, + "pct_cuda_time": 0.8407429228798668, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 44.547, - "cuda_time_us": 17.245, - "pct_cuda_time": 0.3016950380132249, + "cpu_time_us": 49.113, + "cuda_time_us": 17.666, + "pct_cuda_time": 0.3048743657367188, "trace": "" }, "children": [ @@ -16455,8 +16455,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.245, - "pct_cuda_time": 0.3016950380132249, + "cuda_time_us": 17.666, + "pct_cuda_time": 0.3048743657367188, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16466,9 +16466,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 21.902, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06866645544806656, + "cpu_time_us": 23.915, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.0663558211399119, "trace": "" }, "children": [ @@ -16476,8 +16476,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06866645544806656, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.0663558211399119, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -16487,9 +16487,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 135.212, - "cuda_time_us": 12.937, - "pct_cuda_time": 0.22632813608449348, + "cpu_time_us": 171.787, + "cuda_time_us": 13.257, + "pct_cuda_time": 0.2287852069835662, "trace": "" }, "children": [ @@ -16497,8 +16497,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.804, - "pct_cuda_time": 0.04905496587933213, + "cuda_time_us": 2.683, + "pct_cuda_time": 0.04630238442610757, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16507,8 +16507,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.289, - "pct_cuda_time": 0.11002378046188292, + "cuda_time_us": 6.729, + "pct_cuda_time": 0.11612700141754673, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16517,8 +16517,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.844, - "pct_cuda_time": 0.06724938974327843, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.0663558211399119, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16528,9 +16528,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 55.343, - "cuda_time_us": 13.64, - "pct_cuda_time": 0.23862686683098797, + "cpu_time_us": 57.466, + "cuda_time_us": 13.949, + "pct_cuda_time": 0.24072752901966996, "trace": "" }, "children": [ @@ -16538,8 +16538,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.64, - "pct_cuda_time": 0.23862686683098797, + "cuda_time_us": 13.949, + "pct_cuda_time": 0.24072752901966996, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16551,9 +16551,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 24.627, + "cpu_time_us": 27.882, "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "pct_cuda_time": 0.07948892384146533, "trace": "" }, "children": [ @@ -16562,7 +16562,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.606, - "pct_cuda_time": 0.08058030415128523, + "pct_cuda_time": 0.07948892384146533, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16572,18 +16572,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 123.905, - "cuda_time_us": 104.90400000000001, - "pct_cuda_time": 1.8352575394455986, + "cpu_time_us": 151.106, + "cuda_time_us": 105.465, + "pct_cuda_time": 1.8200823606035914, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 47.101, - "cuda_time_us": 62.432, - "pct_cuda_time": 1.0922252602633609, + "cpu_time_us": 59.659, + "cuda_time_us": 62.653, + "pct_cuda_time": 1.0812461019190898, "trace": "" }, "children": [ @@ -16591,8 +16591,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.432, - "pct_cuda_time": 1.0922252602633609, + "cuda_time_us": 62.653, + "pct_cuda_time": 1.0812461019190898, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16602,9 +16602,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 17.356, - "cuda_time_us": 8.772, - "pct_cuda_time": 0.15346296743705473, + "cpu_time_us": 20.952, + "cuda_time_us": 7.931, + "pct_cuda_time": 0.13687074576349578, "trace": "" }, "children": [ @@ -16612,8 +16612,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.772, - "pct_cuda_time": 0.15346296743705473, + "cuda_time_us": 7.931, + "pct_cuda_time": 0.13687074576349578, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -16623,9 +16623,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 44.867, - "cuda_time_us": 33.7, - "pct_cuda_time": 0.5895693117451829, + "cpu_time_us": 54.121, + "cuda_time_us": 34.881, + "pct_cuda_time": 0.6019655129210058, "trace": "" }, "children": [ @@ -16633,8 +16633,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.7, - "pct_cuda_time": 0.5895693117451829, + "cuda_time_us": 34.881, + "pct_cuda_time": 0.6019655129210058, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -16648,18 +16648,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 494.699, - "cuda_time_us": 159.758, - "pct_cuda_time": 2.7949084304387815, + "cpu_time_us": 560.338, + "cuda_time_us": 167.498, + "pct_cuda_time": 2.8906286942244375, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 22.543, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07988051861805652, + "cpu_time_us": 27.08, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "" }, "children": [ @@ -16667,8 +16667,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07988051861805652, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16678,18 +16678,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 299.066, - "cuda_time_us": 46.354, - "pct_cuda_time": 0.8109464651820832, + "cpu_time_us": 335.702, + "cuda_time_us": 50.458, + "pct_cuda_time": 0.870788562568966, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 44.657, - "cuda_time_us": 16.834, - "pct_cuda_time": 0.29450474165929996, + "cpu_time_us": 50.997, + "cuda_time_us": 17.495, + "pct_cuda_time": 0.3019233006093, "trace": "" }, "children": [ @@ -16697,8 +16697,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.834, - "pct_cuda_time": 0.29450474165929996, + "cuda_time_us": 17.495, + "pct_cuda_time": 0.3019233006093, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16708,9 +16708,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 20.781, - "cuda_time_us": 3.444, - "pct_cuda_time": 0.060251534410991396, + "cpu_time_us": 24.626, + "cuda_time_us": 3.484, + "pct_cuda_time": 0.06012579475980573, "trace": "" }, "children": [ @@ -16718,8 +16718,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.444, - "pct_cuda_time": 0.060251534410991396, + "cuda_time_us": 3.484, + "pct_cuda_time": 0.06012579475980573, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -16729,9 +16729,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 148.411, - "cuda_time_us": 12.096, - "pct_cuda_time": 0.21161514524836003, + "cpu_time_us": 170.354, + "cuda_time_us": 15.219, + "pct_cuda_time": 0.26264479634026505, "trace": "" }, "children": [ @@ -16739,8 +16739,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.603, - "pct_cuda_time": 0.0455385435748579, + "cuda_time_us": 2.683, + "pct_cuda_time": 0.04630238442610757, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16749,8 +16749,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.808, - "pct_cuda_time": 0.10160885942480777, + "cuda_time_us": 8.652, + "pct_cuda_time": 0.14931354083290446, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16759,8 +16759,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.685, - "pct_cuda_time": 0.06446774224869432, + "cuda_time_us": 3.884, + "pct_cuda_time": 0.06702887108125301, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16770,9 +16770,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 56.134, - "cuda_time_us": 13.98, - "pct_cuda_time": 0.24457504386343193, + "cpu_time_us": 58.347, + "cuda_time_us": 14.26, + "pct_cuda_time": 0.24609467085959522, "trace": "" }, "children": [ @@ -16780,8 +16780,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.98, - "pct_cuda_time": 0.24457504386343193, + "cuda_time_us": 14.26, + "pct_cuda_time": 0.24609467085959522, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16793,9 +16793,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 25.238, - "cuda_time_us": 4.806, - "pct_cuda_time": 0.08407923181742875, + "cpu_time_us": 30.706, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "" }, "children": [ @@ -16803,8 +16803,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.806, - "pct_cuda_time": 0.08407923181742875, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16814,18 +16814,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 132.578, - "cuda_time_us": 104.03200000000001, - "pct_cuda_time": 1.8200022148212127, + "cpu_time_us": 149.733, + "cuda_time_us": 107.828, + "pct_cuda_time": 1.8608622839725408, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 53.59, - "cuda_time_us": 62.843, - "pct_cuda_time": 1.0994155566172856, + "cpu_time_us": 60.631, + "cuda_time_us": 64.785, + "pct_cuda_time": 1.1180394987124036, "trace": "" }, "children": [ @@ -16833,8 +16833,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.843, - "pct_cuda_time": 1.0994155566172856, + "cuda_time_us": 64.785, + "pct_cuda_time": 1.1180394987124036, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16844,9 +16844,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 17.066, - "cuda_time_us": 8.131, - "pct_cuda_time": 0.14224890426706474, + "cpu_time_us": 19.319, + "cuda_time_us": 8.292, + "pct_cuda_time": 0.14310077214360195, "trace": "" }, "children": [ @@ -16854,8 +16854,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.131, - "pct_cuda_time": 0.14224890426706474, + "cuda_time_us": 8.292, + "pct_cuda_time": 0.14310077214360195, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -16865,9 +16865,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 46.71, - "cuda_time_us": 33.058, - "pct_cuda_time": 0.5783377539368622, + "cpu_time_us": 52.268, + "cuda_time_us": 34.751, + "pct_cuda_time": 0.5997220131165353, "trace": "" }, "children": [ @@ -16875,8 +16875,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.058, - "pct_cuda_time": 0.5783377539368622, + "cuda_time_us": 34.751, + "pct_cuda_time": 0.5997220131165353, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -16890,18 +16890,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 494.399, - "cuda_time_us": 163.732, - "pct_cuda_time": 2.8644321231650527, + "cpu_time_us": 558.385, + "cuda_time_us": 166.357, + "pct_cuda_time": 2.8709376690175095, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 22.964, - "cuda_time_us": 4.605, - "pct_cuda_time": 0.08056280951295451, + "cpu_time_us": 27.31, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "" }, "children": [ @@ -16909,8 +16909,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.605, - "pct_cuda_time": 0.08056280951295451, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08017923147361006, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16920,18 +16920,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 293.989, - "cuda_time_us": 47.865, - "pct_cuda_time": 0.8373808636997975, + "cpu_time_us": 329.894, + "cuda_time_us": 50.278999999999996, + "pct_cuda_time": 0.8676994359151183, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 46.44, - "cuda_time_us": 16.574, - "pct_cuda_time": 0.2899561356933134, + "cpu_time_us": 51.237, + "cuda_time_us": 17.024, + "pct_cuda_time": 0.2937949282407959, "trace": "" }, "children": [ @@ -16939,8 +16939,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.574, - "pct_cuda_time": 0.2899561356933134, + "cuda_time_us": 17.024, + "pct_cuda_time": 0.2937949282407959, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16950,9 +16950,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 22.464, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06866645544806656, + "cpu_time_us": 23.905, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06566551350776717, "trace": "" }, "children": [ @@ -16960,8 +16960,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06866645544806656, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06566551350776717, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -16971,9 +16971,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 140.359, - "cuda_time_us": 13.577, - "pct_cuda_time": 0.23752470461615274, + "cpu_time_us": 163.084, + "cuda_time_us": 14.739, + "pct_cuda_time": 0.25436110475452833, "trace": "" }, "children": [ @@ -16981,8 +16981,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.763, - "pct_cuda_time": 0.04833768570777271, + "cuda_time_us": 2.483, + "pct_cuda_time": 0.042850846265383935, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16991,8 +16991,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.929, - "pct_cuda_time": 0.12122034899354221, + "cuda_time_us": 8.412, + "pct_cuda_time": 0.14517169504003613, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17001,8 +17001,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.06796666991483784, + "cuda_time_us": 3.844, + "pct_cuda_time": 0.06633856344910828, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17012,9 +17012,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 56.996, - "cuda_time_us": 13.789, - "pct_cuda_time": 0.24123356794226486, + "cpu_time_us": 58.728, + "cuda_time_us": 14.711, + "pct_cuda_time": 0.25387788941202705, "trace": "" }, "children": [ @@ -17022,8 +17022,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.789, - "pct_cuda_time": 0.24123356794226486, + "cuda_time_us": 14.711, + "pct_cuda_time": 0.25387788941202705, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17035,9 +17035,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 25.729, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.08267966075097134, + "cpu_time_us": 31.226, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "" }, "children": [ @@ -17045,8 +17045,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.08267966075097134, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17056,18 +17056,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 130.946, - "cuda_time_us": 106.536, - "pct_cuda_time": 1.8638087892013293, + "cpu_time_us": 148.222, + "cuda_time_us": 106.74600000000001, + "pct_cuda_time": 1.8421894625230262, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 52.629, - "cuda_time_us": 62.161, - "pct_cuda_time": 1.0874842132757363, + "cpu_time_us": 59.559, + "cuda_time_us": 64.285, + "pct_cuda_time": 1.1094106533105945, "trace": "" }, "children": [ @@ -17075,8 +17075,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.161, - "pct_cuda_time": 1.0874842132757363, + "cuda_time_us": 64.285, + "pct_cuda_time": 1.1094106533105945, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17086,9 +17086,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 16.826, - "cuda_time_us": 9.213, - "pct_cuda_time": 0.16117810294090118, + "cpu_time_us": 22.133, + "cuda_time_us": 8.251, + "pct_cuda_time": 0.1423932068206536, "trace": "" }, "children": [ @@ -17096,8 +17096,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 9.213, - "pct_cuda_time": 0.16117810294090118, + "cuda_time_us": 8.251, + "pct_cuda_time": 0.1423932068206536, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -17107,9 +17107,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 46.069, - "cuda_time_us": 35.162, - "pct_cuda_time": 0.6151464729846919, + "cpu_time_us": 51.227, + "cuda_time_us": 34.21, + "pct_cuda_time": 0.590385602391778, "trace": "" }, "children": [ @@ -17117,8 +17117,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.162, - "pct_cuda_time": 0.6151464729846919, + "cuda_time_us": 34.21, + "pct_cuda_time": 0.590385602391778, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -17132,18 +17132,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 542.73, - "cuda_time_us": 163.814, - "pct_cuda_time": 2.8658666835081714, + "cpu_time_us": 617.594, + "cuda_time_us": 167.2, + "pct_cuda_time": 2.8854859023649593, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.194, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07988051861805652, + "cpu_time_us": 34.221, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "" }, "children": [ @@ -17151,8 +17151,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07988051861805652, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08086953910575478, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17162,18 +17162,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 325.776, - "cuda_time_us": 49.899, - "pct_cuda_time": 0.872964958064477, + "cpu_time_us": 382.321, + "cuda_time_us": 50.39, + "pct_cuda_time": 0.86961503959432, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 46.739, - "cuda_time_us": 17.395, - "pct_cuda_time": 0.30431923376283254, + "cpu_time_us": 52.168, + "cuda_time_us": 17.145, + "pct_cuda_time": 0.29588310882803365, "trace": "" }, "children": [ @@ -17181,8 +17181,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.395, - "pct_cuda_time": 0.30431923376283254, + "cuda_time_us": 17.145, + "pct_cuda_time": 0.29588310882803365, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17192,9 +17192,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 22.444, + "cpu_time_us": 24.216, "cuda_time_us": 3.685, - "pct_cuda_time": 0.06446774224869432, + "pct_cuda_time": 0.06359459061133298, "trace": "" }, "children": [ @@ -17203,7 +17203,7 @@ "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.685, - "pct_cuda_time": 0.06446774224869432, + "pct_cuda_time": 0.06359459061133298, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -17213,9 +17213,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 170.004, - "cuda_time_us": 14.499, - "pct_cuda_time": 0.25365476115707436, + "cpu_time_us": 210.334, + "cuda_time_us": 15.099, + "pct_cuda_time": 0.2605738734438309, "trace": "" }, "children": [ @@ -17223,8 +17223,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.843, - "pct_cuda_time": 0.04973725677423011, + "cuda_time_us": 2.923, + "pct_cuda_time": 0.050444230218975934, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17233,8 +17233,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.611, - "pct_cuda_time": 0.13315169233509158, + "cuda_time_us": 8.251, + "pct_cuda_time": 0.1423932068206536, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17243,8 +17243,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.045, - "pct_cuda_time": 0.07076581204775266, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06773643640420135, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17254,9 +17254,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 58.297, - "cuda_time_us": 14.32, - "pct_cuda_time": 0.2505232208958759, + "cpu_time_us": 62.293, + "cuda_time_us": 14.461, + "pct_cuda_time": 0.24956346671112248, "trace": "" }, "children": [ @@ -17264,8 +17264,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.32, - "pct_cuda_time": 0.2505232208958759, + "cuda_time_us": 14.461, + "pct_cuda_time": 0.24956346671112248, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17277,9 +17277,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.234, - "cuda_time_us": 4.766, - "pct_cuda_time": 0.08337944628420003, + "cpu_time_us": 28.162, + "cuda_time_us": 4.967, + "pct_cuda_time": 0.08571895022157149, "trace": "" }, "children": [ @@ -17287,8 +17287,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.766, - "pct_cuda_time": 0.08337944628420003, + "cuda_time_us": 4.967, + "pct_cuda_time": 0.08571895022157149, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17298,18 +17298,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 141.872, - "cuda_time_us": 104.583, - "pct_cuda_time": 1.8296417605414381, + "cpu_time_us": 152.128, + "cuda_time_us": 107.157, + "pct_cuda_time": 1.849282373443313, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 53.88, - "cuda_time_us": 62.783, - "pct_cuda_time": 1.0983658783174426, + "cpu_time_us": 60.54, + "cuda_time_us": 65.327, + "pct_cuda_time": 1.1273931671279647, "trace": "" }, "children": [ @@ -17317,8 +17317,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.783, - "pct_cuda_time": 1.0983658783174426, + "cuda_time_us": 65.327, + "pct_cuda_time": 1.1273931671279647, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17328,9 +17328,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 21.462, - "cuda_time_us": 8.531, - "pct_cuda_time": 0.1492467595993518, + "cpu_time_us": 22.754, + "cuda_time_us": 8.371, + "pct_cuda_time": 0.14446412971708777, "trace": "" }, "children": [ @@ -17338,8 +17338,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.531, - "pct_cuda_time": 0.1492467595993518, + "cuda_time_us": 8.371, + "pct_cuda_time": 0.14446412971708777, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -17349,9 +17349,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 51.247, - "cuda_time_us": 33.269, - "pct_cuda_time": 0.5820291226246436, + "cpu_time_us": 51.657, + "cuda_time_us": 33.459, + "pct_cuda_time": 0.5774250765982607, "trace": "" }, "children": [ @@ -17359,8 +17359,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.269, - "pct_cuda_time": 0.5820291226246436, + "cuda_time_us": 33.459, + "pct_cuda_time": 0.5774250765982607, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -17374,18 +17374,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 601.018, - "cuda_time_us": 152.446, - "pct_cuda_time": 2.666987634964574, + "cpu_time_us": 647.879, + "cuda_time_us": 156.115, + "pct_cuda_time": 2.694184399806852, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.431, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cpu_time_us": 27.321, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08294046200218896, "trace": "" }, "children": [ @@ -17393,8 +17393,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08128008968451393, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08294046200218896, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17404,18 +17404,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 386.898, - "cuda_time_us": 48.887, - "pct_cuda_time": 0.8552603840737909, + "cpu_time_us": 419.186, + "cuda_time_us": 49.519999999999996, + "pct_cuda_time": 0.854600848595172, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 50.865, - "cuda_time_us": 17.235, - "pct_cuda_time": 0.3015200916299177, + "cpu_time_us": 49.724, + "cuda_time_us": 17.506, + "pct_cuda_time": 0.3021131352081398, "trace": "" }, "children": [ @@ -17423,8 +17423,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.235, - "pct_cuda_time": 0.3015200916299177, + "cuda_time_us": 17.506, + "pct_cuda_time": 0.3021131352081398, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17434,9 +17434,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.546, - "cuda_time_us": 4.045, - "pct_cuda_time": 0.07076581204775266, + "cpu_time_us": 24.918, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06704612877205661, "trace": "" }, "children": [ @@ -17444,8 +17444,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.045, - "pct_cuda_time": 0.07076581204775266, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06704612877205661, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -17455,9 +17455,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 218.416, - "cuda_time_us": 13.257, - "pct_cuda_time": 0.23192642035032313, + "cpu_time_us": 249.333, + "cuda_time_us": 14.259, + "pct_cuda_time": 0.2460774131687916, "trace": "" }, "children": [ @@ -17465,8 +17465,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.763, - "pct_cuda_time": 0.04833768570777271, + "cuda_time_us": 2.684, + "pct_cuda_time": 0.04631964211691119, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17475,8 +17475,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.609, - "pct_cuda_time": 0.11562206472771257, + "cuda_time_us": 7.65, + "pct_cuda_time": 0.13202133464767907, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17485,8 +17485,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.06796666991483784, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06773643640420135, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 3], float32[1, 32, 3], bfloat16[1, 32, 3, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 33], int32[1], None, 16, 513, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17496,9 +17496,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 57.996, - "cuda_time_us": 14.35, - "pct_cuda_time": 0.25104806004579744, + "cpu_time_us": 60.2, + "cuda_time_us": 13.87, + "pct_cuda_time": 0.2393641714461841, "trace": "" }, "children": [ @@ -17506,8 +17506,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.35, - "pct_cuda_time": 0.25104806004579744, + "cuda_time_us": 13.87, + "pct_cuda_time": 0.2393641714461841, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17519,9 +17519,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.132, - "cuda_time_us": 5.127, - "pct_cuda_time": 0.08969501072158909, + "cpu_time_us": 29.624, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "" }, "children": [ @@ -17529,8 +17529,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.127, - "pct_cuda_time": 0.08969501072158909, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17540,18 +17540,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 142.012, - "cuda_time_us": 93.786, - "pct_cuda_time": 1.64075215048468, + "cpu_time_us": 154.852, + "cuda_time_us": 97.063, + "pct_cuda_time": 1.6750832424715911, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 54.771, - "cuda_time_us": 59.137, - "pct_cuda_time": 1.0345804269636463, + "cpu_time_us": 61.782, + "cuda_time_us": 60.53, + "pct_cuda_time": 1.0446080243430083, "trace": "" }, "children": [ @@ -17559,8 +17559,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 59.137, - "pct_cuda_time": 1.0345804269636463, + "cuda_time_us": 60.53, + "pct_cuda_time": 1.0446080243430083, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17570,9 +17570,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 22.253, - "cuda_time_us": 8.131, - "pct_cuda_time": 0.14224890426706474, + "cpu_time_us": 22.173, + "cuda_time_us": 8.251, + "pct_cuda_time": 0.1423932068206536, "trace": "" }, "children": [ @@ -17580,8 +17580,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.131, - "pct_cuda_time": 0.14224890426706474, + "cuda_time_us": 8.251, + "pct_cuda_time": 0.1423932068206536, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -17591,9 +17591,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 50.295, - "cuda_time_us": 26.518, - "pct_cuda_time": 0.4639228192539691, + "cpu_time_us": 55.433, + "cuda_time_us": 28.282, + "pct_cuda_time": 0.48808201130792933, "trace": "" }, "children": [ @@ -17601,8 +17601,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 26.518, - "pct_cuda_time": 0.4639228192539691, + "cuda_time_us": 28.282, + "pct_cuda_time": 0.48808201130792933, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -17616,9 +17616,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.16, - "cuda_time_us": 4.445, - "pct_cuda_time": 0.0777636673800397, + "cpu_time_us": 28.894, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "" }, "children": [ @@ -17626,8 +17626,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.445, - "pct_cuda_time": 0.0777636673800397, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.07948892384146533, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17639,9 +17639,9 @@ { "entry": { "name": "LogitsProcessor", - "cpu_time_us": 137.114, - "cuda_time_us": 263.732, - "pct_cuda_time": 4.613895956236813, + "cpu_time_us": 146.889, + "cuda_time_us": 261.13800000000003, + "pct_cuda_time": 4.506638861075245, "trace": "" }, "children": [ @@ -17649,8 +17649,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, - "cuda_time_us": 2.884, - "pct_cuda_time": 0.05045453694578954, + "cuda_time_us": 3.204, + "pct_cuda_time": 0.05529364133479265, "trace": "index_select(bfloat16[1, 4096], 0, int64[1])" }, "children": [] @@ -17659,8 +17659,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 260.848, - "pct_cuda_time": 4.563441419291022, + "cuda_time_us": 257.934, + "pct_cuda_time": 4.4513452197404515, "trace": "_rocm_C::wvSplitK(bfloat16[128256, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17670,9 +17670,9 @@ { "entry": { "name": "Sampler", - "cpu_time_us": 1106.702, - "cuda_time_us": 183.09799999999998, - "pct_cuda_time": 3.2032332890777298, + "cpu_time_us": 1013.315, + "cuda_time_us": 186.704, + "pct_cuda_time": 3.2220799037987287, "trace": "" }, "children": [ @@ -17680,8 +17680,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 2.523, - "pct_cuda_time": 0.04413897250840049, + "cuda_time_us": 2.443, + "pct_cuda_time": 0.04216053863323921, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -17690,8 +17690,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 2.522, - "pct_cuda_time": 0.044121477870069765, + "cuda_time_us": 2.482, + "pct_cuda_time": 0.04283358857458032, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -17700,8 +17700,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 2.843, - "pct_cuda_time": 0.04973725677423011, + "cuda_time_us": 2.723, + "pct_cuda_time": 0.0469926920582523, "trace": "copy_(int32[1], int32[1], True) <- _to_copy(int32[1], 3, 0, None, None, True, None) <- to(int32[1], 3, 0, None, None, True, False, None)" }, "children": [] @@ -17710,8 +17710,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 2.443, - "pct_cuda_time": 0.04273940144194308, + "cuda_time_us": 2.523, + "pct_cuda_time": 0.04354115389752867, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -17720,8 +17720,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 2.323, - "pct_cuda_time": 0.040640044842256966, + "cuda_time_us": 2.483, + "pct_cuda_time": 0.042850846265383935, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -17730,8 +17730,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 2.362, - "pct_cuda_time": 0.04132233573715496, + "cuda_time_us": 2.482, + "pct_cuda_time": 0.04283358857458032, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -17740,8 +17740,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 2.403, - "pct_cuda_time": 0.04203961590871438, + "cuda_time_us": 2.483, + "pct_cuda_time": 0.042850846265383935, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -17750,8 +17750,8 @@ "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.244, - "pct_cuda_time": 0.056752606744847875, + "cuda_time_us": 3.444, + "pct_cuda_time": 0.05943548712766101, "trace": "copy_(float32[1, 128256], bfloat16[1, 128256], False) <- _to_copy(bfloat16[1, 128256], 6, None, None, None, False, None) <- to(bfloat16[1, 128256], 6, False, False, None)" }, "children": [] @@ -17760,8 +17760,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2})", "cpu_time_us": 0, - "cuda_time_us": 9.173, - "pct_cuda_time": 0.1604783174076725, + "cuda_time_us": 10.174, + "pct_cuda_time": 0.17557974623601133, "trace": "div_(float32[1, 128256], bfloat16[1, 1])" }, "children": [] @@ -17770,8 +17770,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, - "cuda_time_us": 64.375, - "pct_cuda_time": 1.126217342539945, + "cuda_time_us": 64.495, + "pct_cuda_time": 1.1130347683793544, "trace": "_softmax(float32[1, 128256], -1, False) <- softmax(float32[1, 128256], -1, 6)" }, "children": [] @@ -17780,8 +17780,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, - "cuda_time_us": 35.652, - "pct_cuda_time": 0.6237188457667436, + "cuda_time_us": 35.572, + "pct_cuda_time": 0.6138905772663059, "trace": "_log_softmax(float32[1, 128256], -1, False) <- log_softmax(float32[1, 128256], -1, 6)" }, "children": [] @@ -17790,8 +17790,8 @@ "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 2.202, - "pct_cuda_time": 0.03852319360424014, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.0663558211399119, "trace": "copy_(int64[1], int32[1], False) <- _to_copy(int32[1], 4, None, None, None, False, None) <- to(int32[1], 4, False, False, None)" }, "children": [] @@ -17800,8 +17800,8 @@ "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.528, - "pct_cuda_time": 0.09671036069220684, + "cuda_time_us": 5.808, + "pct_cuda_time": 0.10023266818741437, "trace": "index(float32[1, 128256], None)" }, "children": [] @@ -17810,8 +17810,8 @@ "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4, 4> >(at::native::ReduceOp, unsigned int, long, 4, 4>)", "cpu_time_us": 0, - "cuda_time_us": 40.98, - "pct_cuda_time": 0.7169302787928069, + "cuda_time_us": 41.021, + "pct_cuda_time": 0.7079277344552213, "trace": "argmax(float32[1, 128256], -1, False)" }, "children": [] @@ -17820,8 +17820,8 @@ "entry": { "name": "Memcpy DtoH (Device -> Host)", "cpu_time_us": 0, - "cuda_time_us": 4.525, - "pct_cuda_time": 0.07916323844649711, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08155984673789951, "trace": "copy_(int64[1], int64[1], False) <- _to_copy(int64[1], 4, 0, None, None, False, None) <- to(int64[1], 4, 0, None, None, False, False, None)" }, "children": []