Upload app.py with huggingface_hub

app.py

@@ -41,7 +41,7 @@ _lora_path = hf_hub_download(CMO_LORA, "adapter_model.safetensors", repo_type="m
 _lora_a_pairs = {}  # module_path -> A weight (numpy)
 _lora_b_pairs = {}  # module_path -> B weight (numpy)
 
-with safe_open(_lora_path, framework="numpy", device="cpu") as _f:
+with safe_open(_lora_path, framework="pt", device="cpu") as _f:
     for _key in _f.keys():
         if not _key.startswith("base_model.model."):
             continue
@@ -55,17 +55,15 @@ with safe_open(_lora_path, framework="numpy", device="cpu") as _f:
 
 # Merge LoRA weights into the transformer: w_new = w_orig + scale * (B @ A)
 _merge_count = 0
-for _module_path, _a_np in _lora_a_pairs.items():
+for _module_path, _a_tensor in _lora_a_pairs.items():
     if _module_path not in _lora_b_pairs:
         continue
-    _b_np = _lora_b_pairs[_module_path]
+    _b_tensor = _lora_b_pairs[_module_path]
     # Navigate to the module in the transformer
     _module = pipe.transformer
     for _part in _module_path.split("."):
         _module = getattr(_module, _part)
     # Merge: w_orig + scale * (B @ A)
-    _a_tensor = torch.from_numpy(_a_np)
-    _b_tensor = torch.from_numpy(_b_np)
     _delta = (_b_tensor.float() @ _a_tensor.float()) * LORA_SCALE
     _module.weight.data.add_(_delta.to(_module.weight.data.dtype))
     _merge_count += 1