From 202e61889889a193e438e5072cc4e84bfa5bdbc8 Mon Sep 17 00:00:00 2001
From: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
Date: Tue, 2 Jun 2026 23:27:57 -0700
Subject: [PATCH] Revert "Fix hybrid linear attention misrouting
 plain-RadixAttention linear layers to the full backend (Ring-2.5-1T)"
 (#27116)

---
 .../attention/hybrid_linear_attn_backend.py   | 21 ++++++++++++-------
 .../sglang/srt/models/bailing_moe_linear.py   |  6 ++++++
 2 files changed, 19 insertions(+), 8 deletions(-)

diff --git a/python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py b/python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
index ac5a5e57a..71248dcf2 100644
--- a/python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
+++ b/python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
@@ -782,17 +782,22 @@ class HybridLinearAttnBackend(AttentionBackend):
         self.req_to_token_pool = full_attn_backend.req_to_token_pool
 
     def _is_full_attn(
-        self,
-        layer: Optional[Union[RadixAttention, RadixLinearAttention]],
-        layer_id: Optional[int] = None,
+        self, layer: Optional[RadixAttention], layer_id: Optional[int] = None
     ) -> bool:
-        # RadixLinearAttention is unambiguously a linear-attention layer.
-        # Everything else (including plain RadixAttention) must be classified by
-        # layer id: models like Bailing/Ring use a plain RadixAttention for their
-        # linear layers, so an `isinstance(layer, RadixAttention) -> full` shortcut
-        # would misroute those linear layers to the full-attention backend.
+        # Explicit linear-attention subclass → strong linear signal (KDA, GDN,
+        # Qwen3-Next, Qwen3.5 main linear layers).
         if isinstance(layer, RadixLinearAttention):
             return False
+        # Some hybrid models (Ling-2.5/2.6) wrap their linear layers in plain
+        # `RadixAttention` rather than `RadixLinearAttention`. Those wrappers
+        # set `_is_linear_attention=True` on the attn module so we can
+        # distinguish them from full-attention RadixAttention instances —
+        # including MTP/NEXTN draft layers, which are full and must default to
+        # the full-attn path.
+        if layer is not None and getattr(layer, "_is_linear_attention", False):
+            return False
+        if isinstance(layer, RadixAttention):
+            return True
 
         if layer is not None:
             layer_id = layer.layer_id
diff --git a/python/sglang/srt/models/bailing_moe_linear.py b/python/sglang/srt/models/bailing_moe_linear.py
index c0c6be3ca..1c4a5d872 100644
--- a/python/sglang/srt/models/bailing_moe_linear.py
+++ b/python/sglang/srt/models/bailing_moe_linear.py
@@ -508,6 +508,12 @@ class BailingMoELinearAttention(nn.Module):
             quant_config=quant_config,
             prefix=f"{prefix}.attn",
         )
+        # Marker for HybridLinearAttnBackend._is_full_attn: Bailing wraps
+        # linear-attention layers in a plain RadixAttention, so the
+        # dispatcher can't tell from the type alone that this is a linear
+        # layer (would otherwise default to the full-attn backend, e.g. the
+        # same way MTP/NEXTN draft layers are routed).
+        self.attn._is_linear_attention = True
 
         self.group_norm_size = getattr(config, "group_norm_size", 1)
         self.rms_norm_eps = float(getattr(config, "rms_norm_eps", 1e-5))