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 15, 07:15 UTC
[Misc] Update `WeightsMapper` for qwen2-vl/qwen2.5-vl (#19054)
Signed-off-by: Isotr0py <2037008807@qq.com>
Isotr0py committed
Jun 3, 2025 at 17:08 UTC
ec2dcd80bc173c06a4c48377d4a6b6ca2c78a2f5
2 files changed
+18
-8
vllm/model_executor/models/qwen2_5_vl.py
+9
-4
@@ -823,10 +823,15 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
823
SupportsLoRA, SupportsPP):
824
825
# To ensure correct weight loading and mapping.
826
- hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={
827
- "lm_head.": "language_model.lm_head.",
828
- "model.": "language_model.model.",
829
- })
826
+ hf_to_vllm_mapper = WeightsMapper(
827
+ orig_to_new_prefix={
828
+ # mapping for new names in checkpoint saved after transformers v4.52
829
+ "model.language_model.": "language_model.model.",
830
+ "model.visual.": "visual.",
831
+ # mapping for original checkpoint
832
+ "lm_head.": "language_model.lm_head.",
833
+ "model.": "language_model.model.",
834
+ })
835
836
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
837
super().__init__()
vllm/model_executor/models/qwen2_vl.py
+9
-4
@@ -1071,10 +1071,15 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
1071
SupportsLoRA, SupportsPP):
1072
1073
# To ensure correct weight loading and mapping.
1074
- hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={
1075
- "lm_head.": "language_model.lm_head.",
1076
- "model.": "language_model.model.",
1077
- })
1074
+ hf_to_vllm_mapper = WeightsMapper(
1075
+ orig_to_new_prefix={
1076
+ # mapping for new names in checkpoint saved after transformers v4.52
1077
+ "model.language_model.": "language_model.model.",
1078
+ "model.visual.": "visual.",
1079
+ # mapping for original checkpoint
1080
+ "lm_head.": "language_model.lm_head.",
1081
+ "model.": "language_model.model.",
1082
+ })
1083
1084
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
1085
super().__init__()