[Bugfix] Fix PP for ChatGLM and Molmo (#9422)

2024-10-24 14:12:05 +08:00
parent 056a68c7db
commit 836e8ef6ee
7 changed files with 195 additions and 122 deletions
--- a/vllm/model_executor/models/qwen2_rm.py
+++ b/vllm/model_executor/models/qwen2_rm.py
@@ -119,5 +119,6 @@ class Qwen2ForRewardModel(nn.Module, SupportsPP):
        return self._pooler(hidden_states, pooling_metadata)

    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
-        loader = AutoWeightsLoader(self)
+        loader = AutoWeightsLoader(self,
+                                   ignore_unexpected_prefixes=["lm_head."])
        loader.load_weights(weights)