From d21c31f6814d947fd4edc5cd3ed90ba510b16ffd Mon Sep 17 00:00:00 2001
From: fatSheep <50252876+00fish0@users.noreply.github.com>
Date: Wed, 10 Jun 2026 11:06:50 +0800
Subject: [PATCH] fix: forward update_mamba_state_after_mtp_verify in
 HybridAttnBackend (#25883)

---
 .../sglang/srt/layers/attention/hybrid_attn_backend.py | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/python/sglang/srt/layers/attention/hybrid_attn_backend.py b/python/sglang/srt/layers/attention/hybrid_attn_backend.py
index 92e354dbe..8b280260d 100644
--- a/python/sglang/srt/layers/attention/hybrid_attn_backend.py
+++ b/python/sglang/srt/layers/attention/hybrid_attn_backend.py
@@ -139,6 +139,16 @@ class HybridAttnBackend(AttentionBackend):
         backend = self._select_backend(forward_batch.forward_mode)
         return backend.get_indexer_metadata(layer_id, forward_batch)
 
+    def update_mamba_state_after_mtp_verify(self, *args, **kwargs):
+        # Forward to whichever sub-backend handled target_verify, since its inner
+        # linear_attn_backend.forward_metadata holds the mamba_cache_indices the
+        # method consumes. Mirrors _select_backend's target_verify branch.
+        if self.model_runner.server_args.speculative_attention_mode == "decode":
+            backend = self.decode_backend
+        else:
+            backend = self.prefill_backend
+        return backend.update_mamba_state_after_mtp_verify(*args, **kwargs)
+
     def forward(
         self,
         q: torch.Tensor = None,