fix(qwen3_5): broadcast per-tensor scale in _make_packed_weight_loader for FP8 models (#23062)

2026-04-30 14:16:57 +08:00
parent bcb34da9f9
commit 936c9c2355
2 changed files with 220 additions and 7 deletions
@@ -320,13 +320,10 @@ class Qwen3_5GatedDeltaNet(nn.Module):
                    module, param, loaded_shard_id
                )

-                if len(loaded_weight.shape) == 0:
-                    # Scalar only makes sense for a single logical shard.
-                    assert len(split_sizes) == 1 and split_sizes[0] == 1, (
-                        f"Unexpected scalar for tuple shard load: "
-                        f"{loaded_shard_id=}, {split_sizes=}"
-                    )
-                    chunks = [loaded_weight.reshape(1)]
+                if loaded_weight.numel() == 1:
+                    # Single-element tensor (scalar or [1]):
+                    # broadcast to each logical shard.
+                    chunks = [loaded_weight.view(-1)] * len(loaded_shard_id)
                else:
                    split_dim = getattr(param, "output_dim", 0)
                    if _is_cpu: