From a0757c9624db12767b5abe66d5ec0329873bec7a Mon Sep 17 00:00:00 2001
From: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com>
Date: Mon, 2 Feb 2026 17:06:51 +0800
Subject: [PATCH] [Diffusion] Fix Ring Parallel bug with FA4 (#18062)

---
 python/sglang/jit_kernel/flash_attention/cute/interface.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/python/sglang/jit_kernel/flash_attention/cute/interface.py b/python/sglang/jit_kernel/flash_attention/cute/interface.py
index 10bcd6220..01e53dba1 100644
--- a/python/sglang/jit_kernel/flash_attention/cute/interface.py
+++ b/python/sglang/jit_kernel/flash_attention/cute/interface.py
@@ -1364,6 +1364,7 @@ class FlashAttnVarlenFunc(torch.autograd.Function):
             pack_gqa=pack_gqa,
             score_mod=score_mod,
             aux_tensors=aux_tensors,
+            return_lse=True,
         )
         ctx.save_for_backward(q, k, v, out, lse, cu_seqlens_q, cu_seqlens_k, seqused_q, seqused_k)
         ctx.softmax_scale = softmax_scale