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 6, 21:45 UTC
fix(lora): use replaced_module_name in pooling model name check (#36402)
Signed-off-by: gambletan <ethanchang32@gmail.com> Co-authored-by: Jee Jee Li <pandaleefree@gmail.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Ethan T. committed
Mar 11, 2026 at 18:11 UTC
c87fb515edb180bd66168484e9cae86f384f6215
1 file changed
+3
-3
vllm/lora/model_manager.py
+3
-3
@@ -599,8 +599,8 @@ class LoRAModelManager:
599
replacement_loras[i] = None
600
# HACK Temporary solution for the pool model.
601
if self.is_pooling_model and not lora_model.check_lora_name(module_name):
602
- replaced_module_name = module_name.replace("model.", "")
603
- if lora_model.check_lora_name(module_name):
602
+ replaced_module_name = module_name.removeprefix("model.")
603
+ if lora_model.check_lora_name(replaced_module_name):
604
module_name = replaced_module_name
605
if module_name.endswith(".experts"):
606
if self._is_non_gated_moe and len(replacement_loras) > 0:
@@ -745,7 +745,7 @@ class LoRAModelManager:
745
if self.is_pooling_model and not lora_model.check_lora_name(module_name):
746
# If it's a pool model, and the layer name is not found,
747
# remove the prefix 'model.' and search again.
748
- module_name = module_name.replace("model.", "")
748
+ module_name = module_name.removeprefix("model.")
749
if lora_model.check_lora_name(module_name):
750
org_module_name = module_name
751
logger.info_once(