update

infer_api.py

@@ -279,6 +279,7 @@ def run_multiview_infer(data, pipeline, cfg: TestConfig, num_levels=3):
         generator = torch.Generator(device=pipeline.unet.device).manual_seed(cfg.seed)
 
     if torch.cuda.is_available():
+        pipeline.unet.enable_xformers_memory_efficient_attention()
         pipeline.to(device)
     
     images_cond = []
@@ -344,7 +345,6 @@ def load_multiview_pipeline(cfg):
     pipeline = StableUnCLIPImg2ImgPipeline.from_pretrained(
         cfg.pretrained_path,
         torch_dtype=torch.float16,)
-    pipeline.unet.enable_xformers_memory_efficient_attention()
     return pipeline