[HotFix][Ling 2.6] Fix HybridLinearAttn dispatcher for Ling-2.6 (#26474)

Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>

python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py

@@ -811,9 +811,18 @@ class HybridLinearAttnBackend(AttentionBackend):
     def _is_full_attn(
         self, layer: Optional[RadixAttention], layer_id: Optional[int] = None
     ) -> bool:
-        # Dispatch by the layer's runtime type
+        # 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
 

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))