Upload src/pipeline.py with huggingface_hub

src/pipeline.py

@@ -4,21 +4,23 @@ from typing import TypeAlias
 
 import torch
 from PIL.Image import Image
-from diffusers import FluxPipeline, FluxTransformer2DModel, AutoencoderKL, AutoencoderTiny
+from diffusers import FluxPipeline, FluxTransformer2DModel, AutoencoderKL, AutoencoderTiny, DiffusionPipeline
 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 
+
 
-os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
 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
-torch.backends.cudnn.benchmark = True
-
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
 CHECKPOINT = "jokerbit/flux.1-schnell-Robert-int8wo"
 REVISION = "5ef0012f11a863e5111ec56540302a023bc8587b"
 
@@ -33,7 +35,7 @@ def load_pipeline() -> Pipeline:
             use_safetensors=False,
             local_files_only=True,
             torch_dtype=torch.bfloat16)
-    
+
     pipeline = FluxPipeline.from_pretrained(
         CHECKPOINT,
         revision=REVISION,
@@ -41,9 +43,14 @@ def load_pipeline() -> Pipeline:
         local_files_only=True,
         torch_dtype=torch.bfloat16,
     )
+
     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, fullgraph=True, mode="max-autotune")
     pipeline.to("cuda")
-    for _ in range(4):
+    
+    for _ in range(2):
         pipeline("cat", num_inference_steps=4)
 
     return pipeline
@@ -78,3 +85,4 @@ if __name__ == "__main__":
         infer(request, pipe_)
         stop_time = perf_counter()
         print(f"Request in {stop_time - start_time}s")
+