Upload src/pipeline.py with huggingface_hub

src/pipeline.py

@@ -10,6 +10,7 @@ from pipelines.models import TextToImageRequest
 from torch import Generator
 from torchao.quantization import quantize_, int8_weight_only
 from transformers import T5EncoderModel, CLIPTextModel
+import torch_tensorrt
 
 
 Pipeline: TypeAlias = FluxPipeline
@@ -50,10 +51,10 @@ def load_pipeline() -> Pipeline:
     )
 
     pipeline.transformer.to(memory_format=torch.channels_last)
-    pipeline.text_encoder = torch.compile(pipeline.text_encoder, mode="reduce-overhead")
-#    quantize_(pipeline.vae, int8_weight_only())
-#    pipeline.vae = torch.compile(pipeline.vae, mode="reduce-overhead")
+    #    quantize_(pipeline.vae, int8_weight_only())
+    pipeline.vae = torch.compile(pipeline.vae, backend="tensorrt")
     pipeline.to("cuda")
+    
     for _ in range(2):
         pipeline("cat", num_inference_steps=4)