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 8, 01:15 UTC
[Bug] Fix DeepSeek V3 weight loading caused by incorrect prefix (#34876)
Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Wei Zhao committed
Feb 19, 2026 at 02:20 UTC
7f51e93864709e436fab21c3f4103c49d198f999
1 file changed
+2
-2
vllm/model_executor/models/deepseek_v2.py
+2
-2
@@ -716,7 +716,7 @@ class DeepSeekV2FusedQkvAProj(MergedColumnParallelLinear):
716
def __init__(
717
self,
718
input_size: int,
719
- output_size: int,
719
+ output_size: list[int],
720
quant_config: QuantizationConfig | None = None,
721
prefix: str = "",
722
):
@@ -726,7 +726,7 @@ class DeepSeekV2FusedQkvAProj(MergedColumnParallelLinear):
726
bias=False,
727
quant_config=quant_config,
728
disable_tp=True,
729
- prefix=f"{prefix}.kv_a_proj_with_mqa",
729
+ prefix=prefix,
730
)
731
732
# Check if the DeepSeek V3 fused A GEMM kernel can be used.