Compile

src/pipeline.py

@@ -13,6 +13,7 @@ from transformers import T5EncoderModel, CLIPTextModel
 
 
 Pipeline: TypeAlias = FluxPipeline
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
 torch.backends.cudnn.benchmark = True
 
 CHECKPOINT = "jokerbit/flux.1-schnell-Robert-int8wo"
@@ -37,11 +38,9 @@ def load_pipeline() -> Pipeline:
         local_files_only=True,
         torch_dtype=torch.bfloat16,
     )
-
-    pipeline.to(memory_format=torch.channels_last)
-    pipeline.enable_vae_slicing()
+    pipeline.vae = torch.compile(pipeline.vae, mode="max-autotune")
     pipeline.to("cuda")
-    # quantize_(pipeline.vae, int8_weight_only())
+    
     for _ in range(4):
         pipeline("cat", num_inference_steps=4)