Merge pull request #1 from vinayburugu/vllm_6758

Cherry-pick vllm-project#6758 fix: skip loading lm_head if tie_word_embeddings

vllm/model_executor/models/chameleon.py

@@ -998,6 +998,13 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                 # Models trained using ColossalAI may include these tensors in
                 # the checkpoint. Skip them.
                 continue
+
+            # With tie_word_embeddings, we can skip lm_head.weight
+            # The weight might appear unnecessarily in the files if the model is
+            # processed with quantization, LoRA, fine-tuning, etc.
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
+
             use_default_weight_loading = False
             if "vqmodel" in name:
                 if self.model.vqmodel is not None:

vllm/model_executor/models/llama.py

@@ -469,6 +469,11 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                 # Models trained using ColossalAI may include these tensors in
                 # the checkpoint. Skip them.
                 continue
+            # With tie_word_embeddings, we can skip lm_head.weight
+            # The weight might appear unnecessarily in the files if the model is
+            # processed with quantization, LoRA, fine-tuning, etc.
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
             if scale_name := get_compressed_tensors_cache_scale(name):
                 # Loading kv cache scales for compressed-tensors quantization
                 param = params_dict[scale_name]

vllm/model_executor/models/minicpm.py

@@ -513,7 +513,11 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                 # Models trained using ColossalAI may include these tensors in
                 # the checkpoint. Skip them.
                 continue
-
+            # With tie_word_embeddings, we can skip lm_head.weight
+            # The weight might appear unnecessarily in the files if the model is
+            # processed with quantization, LoRA, fine-tuning, etc.
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
             for (param_name, weight_name, shard_id) in stacked_params_mapping:
                 if weight_name not in name:
                     continue

vllm/model_executor/models/olmo.py

@@ -343,6 +343,11 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                 # Models trained using ColossalAI may include these tensors in
                 # the checkpoint. Skip them.
                 continue
+            # With tie_word_embeddings, we can skip lm_head.weight
+            # The weight might appear unnecessarily in the files if the model is
+            # processed with quantization, LoRA, fine-tuning, etc.
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
             for (param_name, weight_name, shard_id) in stacked_params_mapping:
                 if weight_name not in name:
                     continue