Upload folder using huggingface_hub

app.py

@@ -92,12 +92,24 @@ def load_image_from_pil(pil_image):
 
 # --- Load model at module scope (ZeroGPU pattern) ---------------------------
 print("Loading model...")
+
+# transformers 5.x expects `all_tied_weights_keys` on PreTrainedModel subclasses.
+# The custom InternVLChatModel (written for transformers 4.x) doesn't call post_init()
+# which is where transformers 5.x sets this attribute. Add a default so loading works.
+import transformers.modeling_utils as _mu
+if 'all_tied_weights_keys' not in _mu.PreTrainedModel.__dict__:
+    _mu.PreTrainedModel.all_tied_weights_keys = {}
+
 model = AutoModel.from_pretrained(
     MODEL_ID,
     torch_dtype=torch.bfloat16,
     trust_remote_code=True,
 ).eval().to("cuda")
 
+# Ensure the attribute exists on the loaded model instance (custom code may skip post_init)
+if not hasattr(model, 'all_tied_weights_keys') or not isinstance(getattr(model, 'all_tied_weights_keys', None), dict):
+    model.all_tied_weights_keys = {}
+
 tokenizer = AutoTokenizer.from_pretrained(
     MODEL_ID,
     trust_remote_code=True,

requirements.txt

@@ -1,5 +1,5 @@
 torchvision
 einops
 timm
-transformers==4.56.2
+transformers
 https://huggingface.co/datasets/multimodalart/zerogpu-blackwell-wheels/resolve/main/wheels/pt212-cu130-cp310/flash_attn-2.8.3-cp310-cp310-linux_x86_64.whl