From ca97f7b9bbf2e904065dafc6918af3f9f386fdf0 Mon Sep 17 00:00:00 2001
From: Baekpica <35071468+Baekpica@users.noreply.github.com>
Date: Thu, 30 Apr 2026 16:12:42 +0900
Subject: [PATCH] Fix Gemma4 MoE expert weight remapping (#41206)

Signed-off-by: sunghoon.baek <sunghoon.baek@connectfy.cloud>
Co-authored-by: sunghoon.baek <sunghoon.baek@connectfy.cloud>
Co-authored-by: OpenAI Codex <codex@openai.com>
---
 vllm/model_executor/models/gemma4.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/vllm/model_executor/models/gemma4.py b/vllm/model_executor/models/gemma4.py
index b724fa71968..4c8edf2a01a 100644
--- a/vllm/model_executor/models/gemma4.py
+++ b/vllm/model_executor/models/gemma4.py
@@ -84,6 +84,10 @@ from .utils import (
 logger = init_logger(__name__)
 
 
+def _remap_gemma4_expert_weight_name(name: str) -> str:
+    return re.sub(r"(?<!\.moe)\.experts\.(\d+)\.", r".moe.experts.\1.", name)
+
+
 @triton.jit
 def _gemma4_routing_kernel(
     gating_ptr,
@@ -1650,7 +1654,7 @@ class Gemma4ForCausalLM(
                 # Remap individual 2D expert weights:
                 # .experts.{id}.{proj} → .moe.experts.{id}.{proj}
                 # (This handles per-expert 2D quantized weights)
-                name = re.sub(r"\.experts\.(\d+)\.", r".moe.experts.\1.", name)
+                name = _remap_gemma4_expert_weight_name(name)
 
                 # MoE expert weights: checkpoint stores as 3D packed
                 # tensors.  Explode into per-expert 2D weights for