[BugFix][Pooling] Skip weight-prefix probe when model has WeightsMapper (#50890)

Signed-off-by: Juqi Li <2223621784@qq.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
Li Juqi
2026-08-05 21:12:33 -07:00
committed by GitHub
co-authored by Cyrus Leung
parent f84df12c6a
commit 821717118f
2 changed files with 108 additions and 2 deletions
+13 -1
View File
@@ -175,6 +175,11 @@ def _create_pooling_model_cls(orig_cls: _T) -> _T:
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
params_dict = dict(self.named_parameters())
# Use model-owned remapping only for the probe membership check
# (e.g. Gemma4 model.language_model.* -> language_model.model.*)
# so the "" / "model." scan can early-exit. Forwarded names stay
# unmapped to avoid double-applying the mapper in the parent loader.
hf_to_vllm_mapper = getattr(self, "hf_to_vllm_mapper", None)
# We support loading from both `*ForCausalLM` and `*Model`
candidate_prefixes = ["", "model."]
@@ -185,11 +190,18 @@ def _create_pooling_model_cls(orig_cls: _T) -> _T:
# Clone because the iterator may reuse the tensor buffer
seen_weights.append((name, loaded_weight.clone()))
lookup_name = name
if hf_to_vllm_mapper is not None:
mapped = hf_to_vllm_mapper._map_name(name)
if mapped is None:
continue
lookup_name = mapped
try:
target_prefix = next(
prefix
for prefix in candidate_prefixes
if prefix + name in params_dict
if prefix + lookup_name in params_dict
)
break
except StopIteration: