Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,480 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# FlashAttention
|
2 |
+
This repository provides the official implementation of FlashAttention and
|
3 |
+
FlashAttention-2 from the
|
4 |
+
following papers.
|
5 |
+
|
6 |
+
**FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness**
|
7 |
+
Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, Christopher Ré
|
8 |
+
Paper: https://arxiv.org/abs/2205.14135
|
9 |
+
IEEE Spectrum [article](https://spectrum.ieee.org/mlperf-rankings-2022) about our submission to the MLPerf 2.0 benchmark using FlashAttention.
|
10 |
+

|
11 |
+
|
12 |
+
**FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning**
|
13 |
+
Tri Dao
|
14 |
+
|
15 |
+
Paper: https://tridao.me/publications/flash2/flash2.pdf
|
16 |
+
|
17 |
+

|
18 |
+
|
19 |
+
|
20 |
+
## Usage
|
21 |
+
|
22 |
+
We've been very happy to see FlashAttention being widely adopted in such a short
|
23 |
+
time after its release. This [page](https://github.com/Dao-AILab/flash-attention/blob/main/usage.md)
|
24 |
+
contains a partial list of places where FlashAttention is being used.
|
25 |
+
|
26 |
+
FlashAttention and FlashAttention-2 are free to use and modify (see LICENSE).
|
27 |
+
Please cite and credit FlashAttention if you use it.
|
28 |
+
|
29 |
+
|
30 |
+
## FlashAttention-3 beta release
|
31 |
+
FlashAttention-3 is optimized for Hopper GPUs (e.g. H100).
|
32 |
+
|
33 |
+
Blogpost: https://tridao.me/blog/2024/flash3/
|
34 |
+
|
35 |
+
Paper: https://tridao.me/publications/flash3/flash3.pdf
|
36 |
+
|
37 |
+

|
38 |
+
|
39 |
+
This is a beta release for testing / benchmarking before we integrate that with
|
40 |
+
the rest of the repo.
|
41 |
+
|
42 |
+
Currently released:
|
43 |
+
- FP16 forward and backward
|
44 |
+
|
45 |
+
Coming soon in the next couple of days / next week:
|
46 |
+
- BF16
|
47 |
+
- Variable length (FP16, BF16)
|
48 |
+
- FP8 forward.
|
49 |
+
|
50 |
+
Requirements: H100 / H800 GPU, CUDA >= 12.3.
|
51 |
+
|
52 |
+
To install:
|
53 |
+
```sh
|
54 |
+
cd hopper
|
55 |
+
python setup.py install
|
56 |
+
```
|
57 |
+
To run the test:
|
58 |
+
```sh
|
59 |
+
export PYTHONPATH=$PWD
|
60 |
+
pytest -q -s test_flash_attn.py
|
61 |
+
```
|
62 |
+
|
63 |
+
|
64 |
+
|
65 |
+
## Installation and features
|
66 |
+
|
67 |
+
Requirements:
|
68 |
+
- CUDA 11.6 and above.
|
69 |
+
- PyTorch 1.12 and above.
|
70 |
+
- Linux. Might work for Windows starting v2.3.2 (we've seen a few positive [reports](https://github.com/Dao-AILab/flash-attention/issues/595)) but Windows compilation still requires more testing. If you have ideas on how to set up prebuilt CUDA wheels for Windows, please reach out via Github issue.
|
71 |
+
|
72 |
+
We recommend the
|
73 |
+
[Pytorch](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch)
|
74 |
+
container from Nvidia, which has all the required tools to install FlashAttention.
|
75 |
+
|
76 |
+
To install:
|
77 |
+
1. Make sure that PyTorch is installed.
|
78 |
+
2. Make sure that `packaging` is installed (`pip install packaging`)
|
79 |
+
3. Make sure that `ninja` is installed and that it works correctly (e.g. `ninja
|
80 |
+
--version` then `echo $?` should return exit code 0). If not (sometimes `ninja
|
81 |
+
--version` then `echo $?` returns a nonzero exit code), uninstall then reinstall
|
82 |
+
`ninja` (`pip uninstall -y ninja && pip install ninja`). Without `ninja`,
|
83 |
+
compiling can take a very long time (2h) since it does not use multiple CPU
|
84 |
+
cores. With `ninja` compiling takes 3-5 minutes on a 64-core machine.
|
85 |
+
4. Then:
|
86 |
+
```sh
|
87 |
+
pip install flash-attn --no-build-isolation
|
88 |
+
```
|
89 |
+
Alternatively you can compile from source:
|
90 |
+
```sh
|
91 |
+
python setup.py install
|
92 |
+
```
|
93 |
+
|
94 |
+
If your machine has less than 96GB of RAM and lots of CPU cores, `ninja` might
|
95 |
+
run too many parallel compilation jobs that could exhaust the amount of RAM. To
|
96 |
+
limit the number of parallel compilation jobs, you can set the environment
|
97 |
+
variable `MAX_JOBS`:
|
98 |
+
```sh
|
99 |
+
MAX_JOBS=4 pip install flash-attn --no-build-isolation
|
100 |
+
```
|
101 |
+
|
102 |
+
Interface: `src/flash_attention_interface.py`
|
103 |
+
|
104 |
+
FlashAttention-2 currently supports:
|
105 |
+
1. Ampere, Ada, or Hopper GPUs (e.g., A100, RTX 3090, RTX 4090, H100). Support for Turing
|
106 |
+
GPUs (T4, RTX 2080) is coming soon, please use FlashAttention 1.x for Turing
|
107 |
+
GPUs for now.
|
108 |
+
2. Datatype fp16 and bf16 (bf16 requires Ampere, Ada, or Hopper GPUs).
|
109 |
+
3. All head dimensions up to 256. ~~Head dim > 192 backward requires A100/A800 or H100/H800~~. Head dim 256 backward now works on consumer GPUs (if there's no dropout) as of flash-attn 2.5.5.
|
110 |
+
|
111 |
+
|
112 |
+
## How to use FlashAttention
|
113 |
+
|
114 |
+
The main functions implement scaled dot product attention (softmax(Q @ K^T *
|
115 |
+
softmax_scale) @ V):
|
116 |
+
```python
|
117 |
+
from flash_attn import flash_attn_qkvpacked_func, flash_attn_func
|
118 |
+
```
|
119 |
+
|
120 |
+
```python
|
121 |
+
flash_attn_qkvpacked_func(qkv, dropout_p=0.0, softmax_scale=None, causal=False,
|
122 |
+
window_size=(-1, -1), alibi_slopes=None, deterministic=False):
|
123 |
+
"""dropout_p should be set to 0.0 during evaluation
|
124 |
+
If Q, K, V are already stacked into 1 tensor, this function will be faster than
|
125 |
+
calling flash_attn_func on Q, K, V since the backward pass avoids explicit concatenation
|
126 |
+
of the gradients of Q, K, V.
|
127 |
+
If window_size != (-1, -1), implements sliding window local attention. Query at position i
|
128 |
+
will only attend to keys between [i - window_size[0], i + window_size[1]] inclusive.
|
129 |
+
Arguments:
|
130 |
+
qkv: (batch_size, seqlen, 3, nheads, headdim)
|
131 |
+
dropout_p: float. Dropout probability.
|
132 |
+
softmax_scale: float. The scaling of QK^T before applying softmax.
|
133 |
+
Default to 1 / sqrt(headdim).
|
134 |
+
causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
|
135 |
+
window_size: (left, right). If not (-1, -1), implements sliding window local attention.
|
136 |
+
alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of (-alibi_slope * |i - j|) is added to
|
137 |
+
the attention score of query i and key j.
|
138 |
+
deterministic: bool. Whether to use the deterministic implementation of the backward pass,
|
139 |
+
which is slightly slower and uses more memory. The forward pass is always deterministic.
|
140 |
+
Return:
|
141 |
+
out: (batch_size, seqlen, nheads, headdim).
|
142 |
+
"""
|
143 |
+
```
|
144 |
+
|
145 |
+
```python
|
146 |
+
flash_attn_func(q, k, v, dropout_p=0.0, softmax_scale=None, causal=False,
|
147 |
+
window_size=(-1, -1), alibi_slopes=None, deterministic=False):
|
148 |
+
"""dropout_p should be set to 0.0 during evaluation
|
149 |
+
Supports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads
|
150 |
+
than Q. Note that the number of heads in Q must be divisible by the number of heads in KV.
|
151 |
+
For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head
|
152 |
+
0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.
|
153 |
+
If window_size != (-1, -1), implements sliding window local attention. Query at position i
|
154 |
+
will only attend to keys between
|
155 |
+
[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.
|
156 |
+
|
157 |
+
Arguments:
|
158 |
+
q: (batch_size, seqlen, nheads, headdim)
|
159 |
+
k: (batch_size, seqlen, nheads_k, headdim)
|
160 |
+
v: (batch_size, seqlen, nheads_k, headdim)
|
161 |
+
dropout_p: float. Dropout probability.
|
162 |
+
softmax_scale: float. The scaling of QK^T before applying softmax.
|
163 |
+
Default to 1 / sqrt(headdim).
|
164 |
+
causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
|
165 |
+
window_size: (left, right). If not (-1, -1), implements sliding window local attention.
|
166 |
+
alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of
|
167 |
+
(-alibi_slope * |i + seqlen_k - seqlen_q - j|)
|
168 |
+
is added to the attention score of query i and key j.
|
169 |
+
deterministic: bool. Whether to use the deterministic implementation of the backward pass,
|
170 |
+
which is slightly slower and uses more memory. The forward pass is always deterministic.
|
171 |
+
Return:
|
172 |
+
out: (batch_size, seqlen, nheads, headdim).
|
173 |
+
"""
|
174 |
+
```
|
175 |
+
|
176 |
+
```python
|
177 |
+
def flash_attn_with_kvcache(
|
178 |
+
q,
|
179 |
+
k_cache,
|
180 |
+
v_cache,
|
181 |
+
k=None,
|
182 |
+
v=None,
|
183 |
+
rotary_cos=None,
|
184 |
+
rotary_sin=None,
|
185 |
+
cache_seqlens: Optional[Union[(int, torch.Tensor)]] = None,
|
186 |
+
cache_batch_idx: Optional[torch.Tensor] = None,
|
187 |
+
block_table: Optional[torch.Tensor] = None,
|
188 |
+
softmax_scale=None,
|
189 |
+
causal=False,
|
190 |
+
window_size=(-1, -1), # -1 means infinite context window
|
191 |
+
rotary_interleaved=True,
|
192 |
+
alibi_slopes=None,
|
193 |
+
):
|
194 |
+
"""
|
195 |
+
If k and v are not None, k_cache and v_cache will be updated *inplace* with the new values from
|
196 |
+
k and v. This is useful for incremental decoding: you can pass in the cached keys/values from
|
197 |
+
the previous step, and update them with the new keys/values from the current step, and do
|
198 |
+
attention with the updated cache, all in 1 kernel.
|
199 |
+
|
200 |
+
If you pass in k / v, you must make sure that the cache is large enough to hold the new values.
|
201 |
+
For example, the KV cache could be pre-allocated with the max sequence length, and you can use
|
202 |
+
cache_seqlens to keep track of the current sequence lengths of each sequence in the batch.
|
203 |
+
|
204 |
+
Also apply rotary embedding if rotary_cos and rotary_sin are passed in. The key @k will be
|
205 |
+
rotated by rotary_cos and rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.
|
206 |
+
If causal or local (i.e., window_size != (-1, -1)), the query @q will be rotated by rotary_cos
|
207 |
+
and rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.
|
208 |
+
If not causal and not local, the query @q will be rotated by rotary_cos and rotary_sin at
|
209 |
+
indices cache_seqlens only (i.e. we consider all tokens in @q to be at position cache_seqlens).
|
210 |
+
|
211 |
+
See tests/test_flash_attn.py::test_flash_attn_kvcache for examples of how to use this function.
|
212 |
+
|
213 |
+
Supports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads
|
214 |
+
than Q. Note that the number of heads in Q must be divisible by the number of heads in KV.
|
215 |
+
For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head
|
216 |
+
0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.
|
217 |
+
|
218 |
+
If causal=True, the causal mask is aligned to the bottom right corner of the attention matrix.
|
219 |
+
For example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 = masked out) is:
|
220 |
+
1 1 1 1 0
|
221 |
+
1 1 1 1 1
|
222 |
+
If seqlen_q = 5 and seqlen_k = 2, the causal mask is:
|
223 |
+
0 0
|
224 |
+
0 0
|
225 |
+
0 0
|
226 |
+
1 0
|
227 |
+
1 1
|
228 |
+
If the row of the mask is all zero, the output will be zero.
|
229 |
+
|
230 |
+
If window_size != (-1, -1), implements sliding window local attention. Query at position i
|
231 |
+
will only attend to keys between
|
232 |
+
[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.
|
233 |
+
|
234 |
+
Note: Does not support backward pass.
|
235 |
+
|
236 |
+
Arguments:
|
237 |
+
q: (batch_size, seqlen, nheads, headdim)
|
238 |
+
k_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim) if there's no block_table,
|
239 |
+
or (num_blocks, page_block_size, nheads_k, headdim) if there's a block_table (i.e. paged KV cache)
|
240 |
+
page_block_size must be a multiple of 256.
|
241 |
+
v_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim) if there's no block_table,
|
242 |
+
or (num_blocks, page_block_size, nheads_k, headdim) if there's a block_table (i.e. paged KV cache)
|
243 |
+
k [optional]: (batch_size, seqlen_new, nheads_k, headdim). If not None, we concatenate
|
244 |
+
k with k_cache, starting at the indices specified by cache_seqlens.
|
245 |
+
v [optional]: (batch_size, seqlen_new, nheads_k, headdim). Similar to k.
|
246 |
+
rotary_cos [optional]: (seqlen_ro, rotary_dim / 2). If not None, we apply rotary embedding
|
247 |
+
to k and q. Only applicable if k and v are passed in. rotary_dim must be divisible by 16.
|
248 |
+
rotary_sin [optional]: (seqlen_ro, rotary_dim / 2). Similar to rotary_cos.
|
249 |
+
cache_seqlens: int, or (batch_size,), dtype torch.int32. The sequence lengths of the
|
250 |
+
KV cache.
|
251 |
+
block_table [optional]: (batch_size, max_num_blocks_per_seq), dtype torch.int32.
|
252 |
+
cache_batch_idx: (batch_size,), dtype torch.int32. The indices used to index into the KV cache.
|
253 |
+
If None, we assume that the batch indices are [0, 1, 2, ..., batch_size - 1].
|
254 |
+
If the indices are not distinct, and k and v are provided, the values updated in the cache
|
255 |
+
might come from any of the duplicate indices.
|
256 |
+
softmax_scale: float. The scaling of QK^T before applying softmax.
|
257 |
+
Default to 1 / sqrt(headdim).
|
258 |
+
causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
|
259 |
+
window_size: (left, right). If not (-1, -1), implements sliding window local attention.
|
260 |
+
rotary_interleaved: bool. Only applicable if rotary_cos and rotary_sin are passed in.
|
261 |
+
If True, rotary embedding will combine dimensions 0 & 1, 2 & 3, etc. If False,
|
262 |
+
rotary embedding will combine dimensions 0 & rotary_dim / 2, 1 & rotary_dim / 2 + 1
|
263 |
+
(i.e. GPT-NeoX style).
|
264 |
+
alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of
|
265 |
+
(-alibi_slope * |i + seqlen_k - seqlen_q - j|)
|
266 |
+
is added to the attention score of query i and key j.
|
267 |
+
|
268 |
+
Return:
|
269 |
+
out: (batch_size, seqlen, nheads, headdim).
|
270 |
+
"""
|
271 |
+
```
|
272 |
+
|
273 |
+
To see how these functions are used in a multi-head attention layer (which
|
274 |
+
includes QKV projection, output projection), see the MHA [implementation](https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py).
|
275 |
+
|
276 |
+
## Changelog
|
277 |
+
|
278 |
+
### 2.0: Complete rewrite, 2x faster
|
279 |
+
Upgrading from FlashAttention (1.x) to FlashAttention-2
|
280 |
+
|
281 |
+
These functions have been renamed:
|
282 |
+
- `flash_attn_unpadded_func` -> `flash_attn_varlen_func`
|
283 |
+
- `flash_attn_unpadded_qkvpacked_func` -> `flash_attn_varlen_qkvpacked_func`
|
284 |
+
- `flash_attn_unpadded_kvpacked_func` -> `flash_attn_varlen_kvpacked_func`
|
285 |
+
|
286 |
+
If the inputs have the same sequence lengths in the same batch, it is simpler
|
287 |
+
and faster to use these functions:
|
288 |
+
```python
|
289 |
+
flash_attn_qkvpacked_func(qkv, dropout_p=0.0, softmax_scale=None, causal=False)
|
290 |
+
```
|
291 |
+
```python
|
292 |
+
flash_attn_func(q, k, v, dropout_p=0.0, softmax_scale=None, causal=False)
|
293 |
+
```
|
294 |
+
### 2.1: Change behavior of causal flag
|
295 |
+
|
296 |
+
If seqlen_q != seqlen_k and causal=True, the causal mask is aligned to the
|
297 |
+
bottom right corner of the attention matrix, instead of the top-left corner.
|
298 |
+
|
299 |
+
For example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 =
|
300 |
+
masked out) is:
|
301 |
+
v2.0:
|
302 |
+
1 0 0 0 0
|
303 |
+
1 1 0 0 0
|
304 |
+
v2.1:
|
305 |
+
1 1 1 1 0
|
306 |
+
1 1 1 1 1
|
307 |
+
|
308 |
+
If seqlen_q = 5 and seqlen_k = 2, the causal mask is:
|
309 |
+
v2.0:
|
310 |
+
1 0
|
311 |
+
1 1
|
312 |
+
1 1
|
313 |
+
1 1
|
314 |
+
1 1
|
315 |
+
v2.1:
|
316 |
+
0 0
|
317 |
+
0 0
|
318 |
+
0 0
|
319 |
+
1 0
|
320 |
+
1 1
|
321 |
+
If the row of the mask is all zero, the output will be zero.
|
322 |
+
|
323 |
+
### 2.2: Optimize for inference
|
324 |
+
|
325 |
+
Optimize for inference (iterative decoding) when query has very small sequence
|
326 |
+
length (e.g., query sequence length = 1). The bottleneck here is to load KV
|
327 |
+
cache as fast as possible, and we split the loading across different thread
|
328 |
+
blocks, with a separate kernel to combine results.
|
329 |
+
|
330 |
+
See the function `flash_attn_with_kvcache` with more features for inference
|
331 |
+
(perform rotary embedding, updating KV cache inplace).
|
332 |
+
|
333 |
+
Thanks to the xformers team, and in particular Daniel Haziza, for this
|
334 |
+
collaboration.
|
335 |
+
|
336 |
+
### 2.3: Local (i.e., sliding window) attention
|
337 |
+
|
338 |
+
Implement sliding window attention (i.e., local attention). Thanks to [Mistral
|
339 |
+
AI](https://mistral.ai/) and in particular Timothée Lacroix for this
|
340 |
+
contribution. Sliding window was used in the [Mistral 7B](https://mistral.ai/news/announcing-mistral-7b/) model.
|
341 |
+
|
342 |
+
### 2.4: ALiBi (attention with linear bias), deterministic backward pass.
|
343 |
+
|
344 |
+
Implement ALiBi (Press et al., 2021). Thanks to Sanghun Cho from Kakao Brain for this contribution.
|
345 |
+
|
346 |
+
Implement deterministic backward pass. Thanks to engineers from [Meituan](www.meituan.com) for this contribution.
|
347 |
+
|
348 |
+
### 2.5: Paged KV cache.
|
349 |
+
|
350 |
+
Support paged KV cache (i.e., [PagedAttention](https://arxiv.org/abs/2309.06180)).
|
351 |
+
Thanks to @beginlner for this contribution.
|
352 |
+
|
353 |
+
### 2.6: Softcapping.
|
354 |
+
|
355 |
+
Support attention with softcapping, as used in Gemma-2 and Grok models.
|
356 |
+
Thanks to @Narsil and @lucidrains for this contribution.
|
357 |
+
|
358 |
+
## Performance
|
359 |
+
|
360 |
+
We present expected speedup (combined forward + backward pass) and memory savings from using FlashAttention against PyTorch standard attention, depending on sequence length, on different GPUs (speedup depends on memory bandwidth - we see more speedup on slower GPU memory).
|
361 |
+
|
362 |
+
We currently have benchmarks for these GPUs:
|
363 |
+
* [A100](#a100)
|
364 |
+
* [H100](#h100)
|
365 |
+
<!-- * [RTX 3090](#rtx-3090) -->
|
366 |
+
<!-- * [T4](#t4) -->
|
367 |
+
|
368 |
+
### A100
|
369 |
+
|
370 |
+
We display FlashAttention speedup using these parameters:
|
371 |
+
* Head dimension 64 or 128, hidden dimension 2048 (i.e. either 32 or 16 heads).
|
372 |
+
* Sequence length 512, 1k, 2k, 4k, 8k, 16k.
|
373 |
+
* Batch size set to 16k / seqlen.
|
374 |
+
|
375 |
+
#### Speedup
|
376 |
+
|
377 |
+

|
378 |
+
|
379 |
+
#### Memory
|
380 |
+
|
381 |
+

|
382 |
+
|
383 |
+
We show memory savings in this graph (note that memory footprint is the same no matter if you use dropout or masking).
|
384 |
+
Memory savings are proportional to sequence length -- since standard attention has memory quadratic in sequence length, whereas FlashAttention has memory linear in sequence length.
|
385 |
+
We see 10X memory savings at sequence length 2K, and 20X at 4K.
|
386 |
+
As a result, FlashAttention can scale to much longer sequence lengths.
|
387 |
+
|
388 |
+
### H100
|
389 |
+
|
390 |
+

|
391 |
+
|
392 |
+
## Full model code and training script
|
393 |
+
|
394 |
+
We have released the full GPT model
|
395 |
+
[implementation](https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/models/gpt.py).
|
396 |
+
We also provide optimized implementations of other layers (e.g., MLP, LayerNorm,
|
397 |
+
cross-entropy loss, rotary embedding). Overall this speeds up training by 3-5x
|
398 |
+
compared to the baseline implementation from Huggingface, reaching up to 225
|
399 |
+
TFLOPs/sec per A100, equivalent to 72% model FLOPs utilization (we don't need
|
400 |
+
any activation checkpointing).
|
401 |
+
|
402 |
+
We also include a training
|
403 |
+
[script](https://github.com/Dao-AILab/flash-attention/tree/main/training) to
|
404 |
+
train GPT2 on Openwebtext and GPT3 on The Pile.
|
405 |
+
|
406 |
+
## Triton implementation of FlashAttention
|
407 |
+
|
408 |
+
Phil Tillet (OpenAI) has an experimental implementation of FlashAttention in Triton:
|
409 |
+
https://github.com/openai/triton/blob/master/python/tutorials/06-fused-attention.py
|
410 |
+
|
411 |
+
As Triton is a higher-level language than CUDA, it might be easier to understand
|
412 |
+
and experiment with. The notations in the Triton implementation are also closer
|
413 |
+
to what's used in our paper.
|
414 |
+
|
415 |
+
We also have an experimental implementation in Triton that support attention
|
416 |
+
bias (e.g. ALiBi):
|
417 |
+
https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/flash_attn_triton.py
|
418 |
+
|
419 |
+
|
420 |
+
## Tests
|
421 |
+
We test that FlashAttention produces the same output and gradient as a reference
|
422 |
+
implementation, up to some numerical tolerance. In particular, we check that the
|
423 |
+
maximum numerical error of FlashAttention is at most twice the numerical error
|
424 |
+
of a baseline implementation in Pytorch (for different head dimensions, input
|
425 |
+
dtype, sequence length, causal / non-causal).
|
426 |
+
|
427 |
+
To run the tests:
|
428 |
+
```sh
|
429 |
+
pytest -q -s tests/test_flash_attn.py
|
430 |
+
```
|
431 |
+
## When you encounter issues
|
432 |
+
|
433 |
+
This new release of FlashAttention-2 has been tested on several GPT-style
|
434 |
+
models, mostly on A100 GPUs.
|
435 |
+
|
436 |
+
If you encounter bugs, please open a GitHub Issue!
|
437 |
+
## AMD GPU/ROCm Support
|
438 |
+
ROCm version use [composable_kernel](https://github.com/ROCm/composable_kernel) as backend. It provides the implementation of FlashAttention-2.
|
439 |
+
|
440 |
+
## Installation and features
|
441 |
+
Requirements:
|
442 |
+
- ROCm 6.0+
|
443 |
+
- PyTorch 1.12.1+
|
444 |
+
|
445 |
+
We recommend the
|
446 |
+
[Pytorch](https://hub.docker.com/r/rocm/pytorch)
|
447 |
+
container from ROCm, which has all the required tools to install FlashAttention.
|
448 |
+
|
449 |
+
To compile from source:
|
450 |
+
```sh
|
451 |
+
python setup.py install
|
452 |
+
```
|
453 |
+
|
454 |
+
FlashAttention-2 on ROCm currently supports:
|
455 |
+
1. MI200 or MI300 GPUs.
|
456 |
+
2. Datatype fp16 and bf16
|
457 |
+
3. Forward's head dimensions up to 256. Backward head dimensions up to 128.
|
458 |
+
|
459 |
+
## Tests
|
460 |
+
To run the tests:
|
461 |
+
```sh
|
462 |
+
pytest tests/test_flash_attn_ck.py
|
463 |
+
```
|
464 |
+
|
465 |
+
## Citation
|
466 |
+
If you use this codebase, or otherwise found our work valuable, please cite:
|
467 |
+
```
|
468 |
+
@inproceedings{dao2022flashattention,
|
469 |
+
title={Flash{A}ttention: Fast and Memory-Efficient Exact Attention with {IO}-Awareness},
|
470 |
+
author={Dao, Tri and Fu, Daniel Y. and Ermon, Stefano and Rudra, Atri and R{\'e}, Christopher},
|
471 |
+
booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
|
472 |
+
year={2022}
|
473 |
+
}
|
474 |
+
@inproceedings{dao2023flashattention2,
|
475 |
+
title={Flash{A}ttention-2: Faster Attention with Better Parallelism and Work Partitioning},
|
476 |
+
author={Dao, Tri},
|
477 |
+
booktitle={International Conference on Learning Representations (ICLR)},
|
478 |
+
year={2024}
|
479 |
+
}
|
480 |
+
```
|