Fix model loading

quant_sdxl/quant_sdxl.py

@@ -102,7 +102,10 @@ def main(args):
 
     # Load model from float checkpoint
     print(f"Loading model from {args.model}...")
-    pipe = DiffusionPipeline.from_pretrained(args.model, torch_dtype=dtype)
+    variant = 'fp16' if dtype == torch.float16 else None
+    pipe = DiffusionPipeline.from_pretrained(args.model, torch_dtype=dtype, variant=variant, use_safetensors=True)
+    pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
+    pipe.vae.config.force_upcast=True
     print(f"Model loaded from {args.model}.")
 
     # Move model to target device