From 0af4d764d6626251923aa61adcf16c9bce488454 Mon Sep 17 00:00:00 2001 From: Harry Mellor <19981378+hmellor@users.noreply.github.com> Date: Wed, 26 Mar 2025 17:17:53 +0000 Subject: [PATCH] Fix weight loading for some models in Transformers backend (#15544) Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> --- vllm/model_executor/models/transformers.py | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/vllm/model_executor/models/transformers.py b/vllm/model_executor/models/transformers.py index bdc390689104e..70daadf913798 100644 --- a/vllm/model_executor/models/transformers.py +++ b/vllm/model_executor/models/transformers.py @@ -345,9 +345,11 @@ class TransformersModel(nn.Module): params_dict = dict(self.named_parameters()) loaded_params = set[str]() for name, loaded_weight in weights: - # Necessary for some models which use remote code - if not name.startswith(prefix := self.model.base_model_prefix): - name = maybe_prefix(prefix, name) + # Use "model" instead of base_model_prefix because + # the base model attribute in vLLM is always `model` + if not name.startswith(prefix := "model."): + name = prefix + name + if is_pp_missing_parameter(name, self): continue if name in params_dict: