From ca36d88fa640bcdfc85fec2a267d6a2efa2052a7 Mon Sep 17 00:00:00 2001
From: Yi Zhong <207368749+vincentzed@users.noreply.github.com>
Date: Sun, 8 Feb 2026 09:35:28 -0500
Subject: [PATCH] [ModelOpt] Fix broken Qwen3-235B-A22B-Instruct-2507-NVFP4
 launch (#18189)

Signed-off-by: vincentzed <207368749+vincentzed@users.noreply.github.com>
---
 python/sglang/srt/models/qwen3_moe.py | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/python/sglang/srt/models/qwen3_moe.py b/python/sglang/srt/models/qwen3_moe.py
index 437c4fa3e..3fcf0cfa0 100644
--- a/python/sglang/srt/models/qwen3_moe.py
+++ b/python/sglang/srt/models/qwen3_moe.py
@@ -890,6 +890,14 @@ class Qwen3MoeModel(Qwen2MoeModel):
 class Qwen3MoeForCausalLM(nn.Module):
     fall_back_to_pt_during_load = False
 
+    # Mapping from fused module names to their component weight names.
+    # Required for quantization configs (e.g., ModelOpt FP4) to correctly identify
+    # which layers should be skipped based on the exclude_modules/ignore list.
+    packed_modules_mapping = {
+        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
+        "gate_up_proj": ["gate_proj", "up_proj"],
+    }
+
     def __init__(
         self,
         config: Qwen3MoeConfig,