Upload src/pipeline.py with huggingface_hub

src/pipeline.py

@@ -9,11 +9,17 @@ from huggingface_hub.constants import HF_HUB_CACHE
 from pipelines.models import TextToImageRequest
 from torch import Generator
 from torchao.quantization import quantize_, int8_weight_only
-from transformers import T5EncoderModel, CLIPTextModel
+from transformers import T5EncoderModel, CLIPTextModel, logging 
 
 
 Pipeline: TypeAlias = FluxPipeline
 torch.backends.cudnn.benchmark = True
+torch.backends.cudnn.benchmark = True
+torch._inductor.config.conv_1x1_as_mm = True
+torch._inductor.config.coordinate_descent_tuning = True
+torch._inductor.config.epilogue_fusion = False
+torch._inductor.config.coordinate_descent_check_all_directions = True
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
 
 CHECKPOINT = "jokerbit/flux.1-schnell-Robert-int8wo"
 REVISION = "5ef0012f11a863e5111ec56540302a023bc8587b"
@@ -41,12 +47,14 @@ def load_pipeline() -> Pipeline:
         vae=vae,
         local_files_only=True,
         torch_dtype=torch.bfloat16,
-    ).to("cuda")  
+    )
 
-    pipeline.vae.to(memory_format=torch.channels_last) 
-    quantize_(pipeline.vae, int8_weight_only())
-    pipeline.vae = torch.compile(pipeline.vae, mode="reduce-overhead")
-#    pipeline.to("cuda")
+    pipeline.transformer.to(memory_format=torch.channels_last) 
+    pipeline.vae.to(memory_format=torch.channels_last)
+    quantize_(pipeline.vae, int8_weight_only()) 
+    pipeline.vae = torch.compile(pipeline.vae, mode="reduce-overhead", fullgraph=True)
+    pipeline.to("cuda")
+    
     for _ in range(2):
         pipeline("cat", num_inference_steps=4)