A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 07:15 UTC
[Misc]Remove redundant hidden_size property in ModelConfig (#29749)
Signed-off-by: Xingyu Liu <charlotteliu12x@gmail.com> Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Xingyu Liu committed
Dec 1, 2025 at 01:14 UTC
21c26279344871558aaba033e5bd058ab2e69793
2 files changed
+2
-9
vllm/config/model.py
+1
-8
@@ -1726,19 +1726,12 @@ class ModelConfig:
1726
logger.debug_once("head dtype: %s", head_dtype)
1727
return head_dtype
1728
1729
- @property
1730
- def hidden_size(self):
1731
- if hasattr(self.hf_config, "hidden_size"):
1732
- return self.hf_config.hidden_size
1733
- text_config = self.hf_config.get_text_config()
1734
- return text_config.hidden_size
1735
-
1729
@property
1730
def embedding_size(self):
1731
dense_modules = try_get_dense_modules(self.model, revision=self.revision)
1732
if dense_modules is not None:
1733
return dense_modules[-1]["out_features"]
1741
- return self.hidden_size
1734
+ return self.get_hidden_size()
1735
1736
def get_and_verify_max_len(self, max_model_len: int):
1737
# Consider max_model_len in tokenizer_config only when
vllm/model_executor/models/adapters.py
+1
-1
@@ -301,7 +301,7 @@ def as_seq_cls_model(cls: _T) -> _T:
301
quant_config = vllm_config.quant_config
302
303
self.score = ReplicatedLinear(
304
- model_config.hidden_size,
304
+ model_config.get_hidden_size(),
305
text_config.num_labels,
306
bias=False,
307
params_dtype=vllm_config.model_config.head_dtype,