From f532ad1f9aef2397cc86a00756f0758f8de31956 Mon Sep 17 00:00:00 2001
From: William Hu <67835823+willhu-jpg@users.noreply.github.com>
Date: Mon, 21 Sep 2026 16:52:26 -0400
Subject: [PATCH] Fix GLM-5.3 forget-gate shape for nvCUTEDSL verify (#40607)

---
 python/sglang/srt/layers/attention/linear/kda_backend.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/python/sglang/srt/layers/attention/linear/kda_backend.py b/python/sglang/srt/layers/attention/linear/kda_backend.py
index 54ffc91d5..c526d3432 100644
--- a/python/sglang/srt/layers/attention/linear/kda_backend.py
+++ b/python/sglang/srt/layers/attention/linear/kda_backend.py
@@ -1583,6 +1583,7 @@ class KDAAttnBackend(MambaAttnBackendBase):
             onorm_eps = None
             onorm_gate = None
 
+        a = a.reshape(1, seq_len, h, layer.head_k_dim)
         out = fused_kda_decode_mtp_dspark(
             x_q=x_q,
             x_k=x_k,