Upload handler.py

handler.py

@@ -14,10 +14,8 @@ if IS_COMPILE:
 #from huggingface_inference_toolkit.logging import logger
 
 def compile_pipeline(pipe) -> Any:
-    pipe.transformer.fuse_qkv_projections()
     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.fuse_qkv_projections()
     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
@@ -27,10 +25,12 @@ class EndpointHandler:
         repo_id = "camenduru/FLUX.1-dev-diffusers"
         #repo_id = "NoMoreCopyright/FLUX.1-dev-test"
         dtype = torch.bfloat16
-        quantization_config = TorchAoConfig("int8wo")
+        quantization_config = TorchAoConfig("int8dq")
         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()
         if IS_COMPILE: self.pipeline = compile_pipeline(self.pipeline)
         self.pipeline.to("cuda")
 
@@ -68,3 +68,4 @@ class EndpointHandler:
             output_type="pil",
         ).images[0]
 
+