try cpu inference

leo/inference.py

@@ -166,8 +166,8 @@ def form_batch(data_dict):
 
     
 def inference(scan_id, task, predict_mode=False):
-    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-    # device = 'cpu' # ok for predict_mode=False, and both for Gradio demo local preview
+    # device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+    device = 'cpu' # ok for predict_mode=False, and both for Gradio demo local preview
 
     data_dict = load_data(scan_id)
     data_dict.update(get_lang(task))

leo/model.py

@@ -14,7 +14,7 @@ from leo.grounding_head import SequentialGroundHead
 from leo.utils import get_mlp_head
 
 
-def maybe_autocast(model, dtype='bf16', enabled=True): ### not-half mode
+def maybe_autocast(model, dtype='float32', enabled=True): ### not-half mode
     # if on cpu, don't use autocast
     # if on gpu, use autocast with dtype if provided, otherwise use torch.float16
     enable_autocast = model.device != torch.device('cpu')