Upload handler.py

handler.py

@@ -4,6 +4,8 @@ from typing import Any, Dict
 from diffusers import FluxPipeline, FluxTransformer2DModel, AutoencoderKL, TorchAoConfig
 from PIL import Image
 import torch
+from accelerate import PartialState
+distributed_state = PartialState()
 
 IS_COMPILE = False
 
@@ -14,9 +16,7 @@ if IS_COMPILE:
 #from huggingface_inference_toolkit.logging import logger
 
 def compile_pipeline(pipe) -> Any:
-    pipe.transformer.to(memory_format=torch.channels_last)
     pipe.transformer = torch.compile(pipe.transformer, mode="reduce-overhead", fullgraph=False, dynamic=False, backend="inductor")
-    pipe.vae.to(memory_format=torch.channels_last)
     pipe.vae = torch.compile(pipe.vae, mode="reduce-overhead", fullgraph=False, dynamic=False, backend="inductor")
     return pipe
 
@@ -29,11 +29,12 @@ class EndpointHandler:
         vae = AutoencoderKL.from_pretrained(repo_id, subfolder="vae", torch_dtype=dtype)
         #transformer = FluxTransformer2DModel.from_pretrained(repo_id, subfolder="transformer", torch_dtype=dtype, quantization_config=quantization_config).to("cuda")
         self.pipeline = FluxPipeline.from_pretrained(repo_id, vae=vae, torch_dtype=dtype, quantization_config=quantization_config)
-        self.pipeline.transformer.fuse_qkv_projections()
-        self.pipeline.vae.fuse_qkv_projections()
+        self.pipeline.transformer.fuse_qkv_projections().to(memory_format=torch.channels_last)
+        self.pipeline.vae.fuse_qkv_projections().to(memory_format=torch.channels_last)
         if IS_COMPILE: self.pipeline = compile_pipeline(self.pipeline)
-        self.pipeline.to("cuda")
+        self.pipeline.to(distributed_state.device)
 
+    @torch.inference_mode()
     def __call__(self, data: Dict[str, Any]) -> Image.Image:
         #logger.info(f"Received incoming request with {data=}")