From b4ac2b9c0cba38cbfff6655a55ffd3b656a86b04 Mon Sep 17 00:00:00 2001
From: DarkSharpness <76582120+DarkSharpness@users.noreply.github.com>
Date: Mon, 11 Aug 2025 23:50:23 -0700
Subject: [PATCH] [Fix] Fix dual chunk model default behavior (#9032)

---
 .../attention/dual_chunk_flashattention_backend.py  |  2 +-
 python/sglang/srt/model_executor/model_runner.py    | 13 +++++++++++++
 2 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/python/sglang/srt/layers/attention/dual_chunk_flashattention_backend.py b/python/sglang/srt/layers/attention/dual_chunk_flashattention_backend.py
index ea97ada22..84876b438 100644
--- a/python/sglang/srt/layers/attention/dual_chunk_flashattention_backend.py
+++ b/python/sglang/srt/layers/attention/dual_chunk_flashattention_backend.py
@@ -483,7 +483,7 @@ class DualChunkFlashAttentionBackend(AttentionBackend):
         ).squeeze(1)
         return o.view(-1, layer.tp_q_head_num * layer.v_head_dim)
 
-    def init_cuda_graph_state(self, max_bs: int):
+    def init_cuda_graph_state(self, max_bs: int, max_num_tokens: int):
         """Initialize CUDA graph state for the attention backend.
 
         Args:
diff --git a/python/sglang/srt/model_executor/model_runner.py b/python/sglang/srt/model_executor/model_runner.py
index 7681d5fe0..c865daf6b 100644
--- a/python/sglang/srt/model_executor/model_runner.py
+++ b/python/sglang/srt/model_executor/model_runner.py
@@ -388,6 +388,19 @@ class ModelRunner:
         ):  # override the default attention backend
             server_args.attention_backend = server_args.prefill_attention_backend
 
+        if (
+            getattr(self.model_config.hf_config, "dual_chunk_attention_config", None)
+            is not None
+        ):
+            if server_args.attention_backend is None:
+                server_args.attention_backend = "dual_chunk_flash_attn"
+                logger.info("Dual chunk attention is turned on by default.")
+            elif server_args.attention_backend != "dual_chunk_flash_attn":
+                raise ValueError(
+                    "Dual chunk attention is enabled, but attention backend is set to "
+                    f"{server_args.attention_backend}. Please set it to 'dual_chunk_flash_attn'."
+                )
+
         if server_args.attention_backend is None:
             """
             Auto select the fastest attention backend.