Update app.py

app.py

@@ -127,20 +127,21 @@ def scheduler_swap_callback(pipeline, step_index, timestep, callback_kwargs):
     return {"latents": callback_kwargs["latents"]}   
 
 def load_and_prepare_model():
+    sched = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler',beta_schedule="scaled_linear", beta_start=0.00085, beta_end=0.012, steps_offset=1) #,use_karras_sigmas=True)
     vaeXL = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae", safety_checker=None, use_safetensors=False).to(torch.bfloat16) #.to(device=device, dtype=torch.bfloat16)
     #vaeRV = AutoencoderKL.from_pretrained("SG161222/RealVisXL_V5.0", subfolder='vae', safety_checker=None, use_safetensors=False).to(device).to(torch.bfloat16) #.to(device=device, dtype=torch.bfloat16)
     #sched = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler',beta_schedule="scaled_linear")
-    sched = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler')
+    #sched = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler')
     pipe = StableDiffusionXLPipeline.from_pretrained(
         'ford442/RealVisXL_V5.0_BF16',
         #torch_dtype=torch.bfloat16,
         add_watermarker=False,
       #  low_cpu_mem_usage = False,
         token = HF_TOKEN,
-        scheduler = sched,
+       # scheduler = sched,
     )
     pipe.vae = vaeXL #.to(torch.bfloat16)
-    #pipe.scheduler = sched,
+    pipe.scheduler = sched,
     #pipe.vae.do_resize=False
     #pipe.vae.vae_scale_factor=8
    # pipe.to(device=device, dtype=torch.bfloat16)
@@ -161,20 +162,20 @@ def load_and_prepare_model():
     '''
 
     pipe.unet = pipe.unet.to(memory_format=torch.contiguous_format)
+    pipe.unet = torch.compile(pipe.unet, backend="hidet")
 
     return pipe
     
 hidet.option.parallel_build(True)
-hidet.option.parallel_tune(-1,4.0)
+hidet.option.parallel_tune(-1,8.0)
 
 torch._dynamo.config.suppress_errors = True
 torch._dynamo.disallow_in_graph(diffusers.models.attention.BasicTransformerBlock)
 # Preload and compile both models
-pipe = load_and_prepare_model()
 
 
 # more search
-hidet.torch.dynamo_config.search_space(0)
+hidet.torch.dynamo_config.search_space(1)
 #hidet.torch.dynamo_config.dump_graph_ir("./local_graph")
 hidet.option.cache_dir("local_cache")
 # automatically transform the model to use float16 data type
@@ -183,8 +184,8 @@ hidet.option.cache_dir("local_cache")
 #hidet.torch.dynamo_config.use_fp16_reduction(True)
 # use tensorcore
 hidet.torch.dynamo_config.use_tensor_core()
-pipe.unet = torch.compile(pipe.unet, backend="hidet")
 
+pipe = load_and_prepare_model()
 
 MAX_SEED = np.iinfo(np.int32).max