From 4637970dfb369d900936136c98cb5a6663385884 Mon Sep 17 00:00:00 2001
From: hlu1 <14827759+hlu1@users.noreply.github.com>
Date: Fri, 6 Feb 2026 19:59:42 -0800
Subject: [PATCH] [Qwen3Next] Optimize
 fused_sigmoid_gating_delta_rule_update_kernel (#18271)

---
 .../srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py  | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/python/sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py b/python/sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py
index 5d4722904..a38f9ce50 100644
--- a/python/sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py
+++ b/python/sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py
@@ -190,7 +190,7 @@ def fused_sigmoid_gating_delta_rule_update(
     B, T, H, K, V = *k.shape, v.shape[-1]
     HV = v.shape[2]
     N = B if cu_seqlens is None else len(cu_seqlens) - 1
-    BK, BV = triton.next_power_of_2(K), min(triton.next_power_of_2(V), 8)
+    BK, BV = triton.next_power_of_2(K), min(triton.next_power_of_2(V), 32)
     NK, NV = triton.cdiv(K, BK), triton.cdiv(V, BV)
     assert NK == 1, "NK > 1 is not supported yet"
     num_stages = 3