Upload src/pipeline.py with huggingface_hub

src/pipeline.py

@@ -13,12 +13,14 @@ 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
+
+#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"
 os.environ["TOKENIZERS_PARALLELISM"] = "True"
 CHECKPOINT = "jokerbit/flux.1-schnell-Robert-int8wo"
@@ -42,16 +44,18 @@ def load_pipeline() -> Pipeline:
         transformer=transformer,
         local_files_only=True,
         torch_dtype=torch.bfloat16,
-    )
+    ).to("cuda")
 
     pipeline.transformer.to(memory_format=torch.channels_last)
+    pipeline.transformer = torch.compile(pipeline.transformer, fullgraph=True)
     # 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(2):
-        pipeline("cat", num_inference_steps=4)
+    # pipeline.vae = torch.compile(pipeline.vae, fullgraph=True, mode="max-autotune")
+    # pipeline.to("cuda")
+    PROMPT = 'semiconformity, peregrination, quip, twineless, emotionless, tawa, depickle'
+  
+    for _ in range(4):
+        pipeline(PROMPT, max_sequence_length=256, height=1024, width=1024, num_inference_steps=4, guidance_scale=0.0)
 
     return pipeline
 
@@ -76,13 +80,14 @@ if __name__ == "__main__":
             height=None,
             width=None,
             seed=666)
+    generator = torch.Generator(device="cuda")
     start_time = perf_counter()
     pipe_ = load_pipeline()
     stop_time = perf_counter()
     print(f"Pipeline is loaded in {stop_time - start_time}s")
     for _ in range(4):
         start_time = perf_counter()
-        infer(request, pipe_)
+        infer(request, pipe_, generator=generator.manual_seed(request.seed))
         stop_time = perf_counter()
         print(f"Request in {stop_time - start_time}s")