From 32a1d55431c6a68ab2b599f857c310ba757497f5 Mon Sep 17 00:00:00 2001
From: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>
Date: Sat, 5 Sep 2026 13:42:29 +0800
Subject: [PATCH] fix(modelopt_fp4): skip NVFP4 swiglu-fusion interleave for
 shared experts with swiglu_limit (#37378)

---
 python/sglang/srt/models/deepseek_v2.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py
index 9aafb6edd..45a84a6fb 100644
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -787,6 +787,7 @@ class DeepseekV2MoE(nn.Module):
                     ModelOptFp4LinearMethod,
                 )
                 and fc1_n % 128 == 0
+                and self.shared_experts.swiglu_limit is None
                 and not check_cuda_graph_backend(Phase.PREFILL, Backend.TC_PIECEWISE)
             ):
                 self.shared_experts.gate_up_proj._interleave_for_swiglu_fusion = True