From 295f85df08b14c5d83c54ebb4e94e9754b9b65af Mon Sep 17 00:00:00 2001
From: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
Date: Fri, 10 Jul 2026 08:40:12 +0800
Subject: [PATCH] Fix DFlash mamba verify init ordering (#30680)

---
 python/sglang/srt/speculative/dflash_worker_v2.py | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)

diff --git a/python/sglang/srt/speculative/dflash_worker_v2.py b/python/sglang/srt/speculative/dflash_worker_v2.py
index 28c05ecb7..a35363658 100644
--- a/python/sglang/srt/speculative/dflash_worker_v2.py
+++ b/python/sglang/srt/speculative/dflash_worker_v2.py
@@ -118,12 +118,7 @@ class DFlashWorkerV2(BaseSpecWorker):
         self.nccl_port = nccl_port
         self._target_worker = target_worker
         self.model_runner = target_worker.model_runner
-        self._need_mamba_verify_commit = (
-            self.model_runner.mambaish_config is not None
-            and hasattr(
-                self.model_runner.attn_backend, "update_mamba_state_after_mtp_verify"
-            )
-        )
+        self._need_mamba_verify_commit = False
         self.page_size = server_args.page_size
         # Normalized in arg_groups.speculative_hook.handle_speculative_decoding.
         self.draft_window_size: Optional[int] = (
@@ -286,6 +281,13 @@ class DFlashWorkerV2(BaseSpecWorker):
 
     def init_attention_backends(self):
         self._draft_worker.init_attention_backends()
+        self._need_mamba_verify_commit = (
+            self.model_runner.mambaish_config is not None
+            and hasattr(
+                self.model_runner.attn_backend,
+                "update_mamba_state_after_mtp_verify",
+            )
+        )
 
     def init_cuda_graphs(self):
         capture_decode_cuda_graph = (