From 4673ca1d7869cafa469e327dbe079eefebd8e368 Mon Sep 17 00:00:00 2001
From: Yufeng He <40085740+he-yufeng@users.noreply.github.com>
Date: Wed, 10 Jun 2026 23:47:04 +0800
Subject: [PATCH] fix: prefix DeepSeek V4 MTP projections (#44821)

Signed-off-by: Yufeng He <40085740+he-yufeng@users.noreply.github.com>
---
 vllm/models/deepseek_v4/amd/mtp.py    | 2 ++
 vllm/models/deepseek_v4/nvidia/mtp.py | 2 ++
 2 files changed, 4 insertions(+)

diff --git a/vllm/models/deepseek_v4/amd/mtp.py b/vllm/models/deepseek_v4/amd/mtp.py
index 9317b72144b..37ce8074af4 100644
--- a/vllm/models/deepseek_v4/amd/mtp.py
+++ b/vllm/models/deepseek_v4/amd/mtp.py
@@ -86,6 +86,7 @@ class DeepSeekV4MultiTokenPredictorLayer(nn.Module):
             bias=False,
             return_bias=False,
             quant_config=quant_config,
+            prefix=f"{prefix}.e_proj",
         )
         self.h_proj = ReplicatedLinear(
             config.hidden_size,
@@ -93,6 +94,7 @@ class DeepSeekV4MultiTokenPredictorLayer(nn.Module):
             bias=False,
             return_bias=False,
             quant_config=quant_config,
+            prefix=f"{prefix}.h_proj",
         )
 
         self.hc_eps = config.hc_eps
diff --git a/vllm/models/deepseek_v4/nvidia/mtp.py b/vllm/models/deepseek_v4/nvidia/mtp.py
index 3bf7c6a233f..64715deae99 100644
--- a/vllm/models/deepseek_v4/nvidia/mtp.py
+++ b/vllm/models/deepseek_v4/nvidia/mtp.py
@@ -92,6 +92,7 @@ class DeepSeekV4MultiTokenPredictorLayer(nn.Module):
             bias=False,
             return_bias=False,
             quant_config=quant_config,
+            prefix=f"{prefix}.e_proj",
         )
         self.h_proj = ReplicatedLinear(
             config.hidden_size,
@@ -99,6 +100,7 @@ class DeepSeekV4MultiTokenPredictorLayer(nn.Module):
             bias=False,
             return_bias=False,
             quant_config=quant_config,
+            prefix=f"{prefix}.h_proj",
         )
 
         self.hc_eps = config.hc_eps