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, 01:00 UTC
[BUGFIX] deepseek-v2-lite failed due to fused_qkv_a_proj name update (#21414)
Signed-off-by: Chendi.Xue <chendi.xue@intel.com>
Chendi.Xue committed
Jul 22, 2025 at 22:33 UTC
08d2bd78dafd992bf3c687852084ce8df7109828
1 file changed
+5
-2
vllm/model_executor/models/deepseek_v2.py
+5
-2
@@ -885,13 +885,16 @@ class DeepseekV2ForCausalLM(nn.Module, SupportsPP, MixtureOfExperts):
885
# for mlp.experts[0].gate_gate_up_proj, which breaks load.
886
if (("mlp.experts." in name) and name not in params_dict):
887
continue
888
- name = name.replace(weight_name, param_name)
888
+ name_mapped = name.replace(weight_name, param_name)
889
890
# QKV fusion is optional, fall back to normal
891
# weight loading if it's not enabled
892
+ # if go with fusion option, then update name
893
if ((param_name == "fused_qkv_a_proj")
893
- and name not in params_dict):
894
+ and name_mapped not in params_dict):
895
continue
896
+ else:
897
+ name = name_mapped
898
# Skip loading extra bias for GPTQ models.
899
if name.endswith(".bias") and name not in params_dict:
900
continue