Upload 2 files

handler.py

@@ -15,10 +15,10 @@ import subprocess
 subprocess.run("pip list", shell=True)
 
 IS_NEW_GPU = False
-IS_COMPILE = True
-IS_TURBO = False
+IS_COMPILE = False
+IS_TURBO = True
 IS_4BIT = True
-IS_AUTOQ = True
+IS_AUTOQ = False
 
 # Set high precision for float32 matrix multiplications.
 # This setting optimizes performance on NVIDIA GPUs with Ampere architecture (e.g., A100, RTX 30 series) or newer.

requirements.txt

@@ -4,8 +4,7 @@ torchvision
 torchaudio
 huggingface_hub
 torchao>=0.9.0
-#diffusers==0.32.2
-git+https://github.com/huggingface/diffusers
+diffusers==0.32.2
 peft
 transformers==4.48.3
 numpy
@@ -13,4 +12,5 @@ scipy
 Pillow
 sentencepiece
 protobuf
-triton
+triton
+gemlite