[Bugfix] Fix default weight loading for scalars (vllm-project#7534)

vllm/model_executor/model_loader/weight_utils.py

@@ -516,11 +516,17 @@ def default_weight_loader(param: torch.Tensor,
                           loaded_weight: torch.Tensor) -> None:
     """Default weight loader."""
     try:
-        assert param.size() == loaded_weight.size(), (
-            f"Attempted to load weight ({loaded_weight.size()}) "
-            f"into parameter ({param.size()})")
-
-        param.data.copy_(loaded_weight)
+        if param.numel() == 1 and loaded_weight.numel() == 1:
+            # Sometimes scalar values aren't considered tensors with shapes
+            # so if both param and loaded_weight are a scalar,
+            # "broadcast" instead of copy
+            param.data.fill_(loaded_weight.item())
+        else:
+            assert param.size() == loaded_weight.size(), (
+                f"Attempted to load weight ({loaded_weight.size()}) "
+                f"into parameter ({param.size()})")
+
+            param.data.copy_(loaded_weight)
     except Exception:
         # NOTE: This exception is added for the purpose of setting breakpoint to
         # debug weight loading issues.