From 688a6d23f144369a7d4d466addcba358022f5d33 Mon Sep 17 00:00:00 2001
From: Alison Shao <54658187+alisonshao@users.noreply.github.com>
Date: Sun, 19 Jul 2026 14:42:04 -0700
Subject: [PATCH] [DeepSeek-V4] Fix idle-rank dummy-extend sparse-prefill crash
 under DP breakable CUDA graph (#31705)

---
 python/sglang/srt/layers/attention/deepseek_v4_backend.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/python/sglang/srt/layers/attention/deepseek_v4_backend.py b/python/sglang/srt/layers/attention/deepseek_v4_backend.py
index cf5fa0183..5d53b46ad 100644
--- a/python/sglang/srt/layers/attention/deepseek_v4_backend.py
+++ b/python/sglang/srt/layers/attention/deepseek_v4_backend.py
@@ -94,6 +94,8 @@ def _get_logical_forward_mode(forward_batch: ForwardBatch) -> ForwardMode:
     # from a reused/padded ForwardBatch turn an empty rank into TARGET_VERIFY.
     if forward_batch.forward_mode.is_idle():
         return forward_batch.forward_mode
+    if forward_batch.forward_mode == ForwardMode.EXTEND:
+        return forward_batch.forward_mode
     return (
         getattr(forward_batch, "_original_forward_mode", None)
         or forward_batch.forward_mode