'update'

app.py

@@ -165,6 +165,9 @@ def prepare_data(single_image, crop_size, cfg):
 scene = 'scene'
 @spaces.GPU
 def run_pipeline(pipeline, cfg, single_image, guidance_scale, steps, seed, crop_size, chk_group=None):
+    pipeline.to(device=f'cuda:{_GPU_ID}')
+    pipeline.unet.enable_xformers_memory_efficient_attention()
+    
     global scene
     # pdb.set_trace()
 
@@ -186,6 +189,10 @@ def run_pipeline(pipeline, cfg, single_image, guidance_scale, steps, seed, crop_
     prompt_embeddings = torch.cat([normal_prompt_embeddings, clr_prompt_embeddings], dim=0)
     prompt_embeddings = rearrange(prompt_embeddings, "B Nv N C -> (B Nv) N C")
     
+    
+    imgs_in = imgs_in.to(device=f'cuda:{_GPU_ID}', dtype=weight_dtype)
+    prompt_embeddings = prompt_embeddings.to(device=f'cuda:{_GPU_ID}', dtype=weight_dtype)
+    
     ic(pipeline.unet.device)
     ic(pipeline.unet.dtype)
     
@@ -304,8 +311,7 @@ def run_demo():
 
     pipeline = load_era3d_pipeline(cfg)
     torch.set_grad_enabled(False)
-    pipeline.to(device=f'cuda:{_GPU_ID}')
-    # pipeline.unet.enable_xformers_memory_efficient_attention()
+
     
     predictor = sam_init()