Update app.py

app.py

@@ -15,6 +15,7 @@ from qwen_vl_utils import process_vision_info
 from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration, TextIteratorStreamer
 
 DEFAULT_CKPT_PATH = 'farrell236/OpthModel32B_b'
+DEFAULT_REVISION = '7d10823654d2d3ba42d74d8177cd54368bc7df96'
 # DEFAULT_CKPT_PATH = '/scratch/llm-weights/Qwen/Qwen2.5-VL-7B-Instruct'
 AUTH_TOKEN = os.environ.get("HF_spaces")
 
@@ -25,6 +26,11 @@ def _get_args():
                         type=str,
                         default=DEFAULT_CKPT_PATH,
                         help='Checkpoint name or path, default to %(default)r')
+    parser.add_argument('-r',
+                        '--revision',
+                        type=str,
+                        default=DEFAULT_REVISION,
+                        help='Commit tag of checkpoint, default to %(default)r')
     parser.add_argument('-t',
                         '--auth-token',
                         type=str,
@@ -60,6 +66,7 @@ def _load_model_processor(args):
     if args.flash_attn2:
         model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
             args.checkpoint_path,
+            revision=args.revision,
             use_auth_token=args.auth_token,
             torch_dtype=torch.bfloat16,
             attn_implementation='flash_attention_2',
@@ -67,6 +74,7 @@ def _load_model_processor(args):
     else:
         model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
             args.checkpoint_path, 
+            revision=args.revision,
             use_auth_token=args.auth_token,
             torch_dtype=torch.bfloat16,
             device_map=device_map)