Update app.py

app.py

@@ -20,7 +20,7 @@ os.environ["PATH"] = conda_bin + os.pathsep + os.environ["PATH"]
 # Activate the base environment (adjust if needed)
 os.system(f'{conda_bin}/conda init  --all') 
 os.system(f'{conda_bin}/conda activate base') 
-#os.system(f'{conda_bin}/conda install nvidia/label/cudnn-9.3.0::cudnn')
+os.system(f'{conda_bin}/conda install nvidia/label/cudnn-9.3.0::cudnn')
 os.system(f'{conda_bin}/conda install nvidia/label/cuda-12.4.0::cuda-libraries')
 os.system(f'{conda_bin}/conda install nvidia/label/cuda-12.4.0::cuda-libraries-dev')
 os.system(f'{conda_bin}/conda install nvidia/label/cuda-12.4.0::cuda-cudart')
@@ -34,6 +34,7 @@ os.system(f'{conda_bin}/conda install nvidia/label/cuda-12.4.0::cuda-nvcc')
 #subprocess.run(['sh', './hidet.sh'])
 import hidet
 #print(dir(hidet)) 
+import torch_tensorrt
 
 import random
 import uuid
@@ -194,8 +195,9 @@ def load_and_prepare_model():
     
     #pipe.unet.to(memory_format=torch.channels_last)
     #pipe.enable_vae_tiling()
-    pipe.unet = torch.compile(pipe.unet, backend="hidet")
+    #pipe.unet = torch.compile(pipe.unet, backend="hidet")
     #pipe.unet = torch.compile(pipe.unet, backend="cudagraphs", dynamic=False)
+    pipe.unet = torch.compile(pipe.unet, backend="torch_tensorrt", dynamic=False, options={"precision": torch.bfloat16,"optimization_level": 4,})
     pipe.to(device=device, dtype=torch.bfloat16)
 
     return pipe