todo for layer-independent parameters in get_allow_patterns

2025-11-27 19:26:02 +00:00
parent 20d73e90cd
commit b43d30563d
1 changed files with 3 additions and 0 deletions
@@ -105,6 +105,9 @@ def get_allow_patterns(weight_map: dict[str, str], shard: ShardMetadata) -> list
            ):
                shard_specific_patterns.add(filename)
        sorted_file_names = sorted(weight_map.values())
+        # TODO: if the model needs any "layer-independent" parameters,
+        # we might want to always add files that correspond to them
+        # e.g. lm_head
        if shard.is_first_layer:
            shard_specific_patterns.add(sorted_file_names[0])
        elif shard.is_last_layer: