From 7f51e93864709e436fab21c3f4103c49d198f999 Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Thu, 19 Feb 2026 02:20:30 -0500
Subject: [PATCH] [Bug] Fix DeepSeek V3 weight loading caused by incorrect
 prefix (#34876)

Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
---
 vllm/model_executor/models/deepseek_v2.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/vllm/model_executor/models/deepseek_v2.py b/vllm/model_executor/models/deepseek_v2.py
index 6ed7505c9..3b3b7a1a3 100644
--- a/vllm/model_executor/models/deepseek_v2.py
+++ b/vllm/model_executor/models/deepseek_v2.py
@@ -716,7 +716,7 @@ class DeepSeekV2FusedQkvAProj(MergedColumnParallelLinear):
     def __init__(
         self,
         input_size: int,
-        output_size: int,
+        output_size: list[int],
         quant_config: QuantizationConfig | None = None,
         prefix: str = "",
     ):
@@ -726,7 +726,7 @@ class DeepSeekV2FusedQkvAProj(MergedColumnParallelLinear):
             bias=False,
             quant_config=quant_config,
             disable_tp=True,
-            prefix=f"{prefix}.kv_a_proj_with_mqa",
+            prefix=prefix,
         )
 
         # Check if the DeepSeek V3 fused A GEMM kernel can be used.