[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))

test/registered/8-gpu-models/test_ling_2_6_flash.py

@@ -0,0 +1,54 @@
+"""GSM8K accuracy test for Ling-2.6-flash (BailingMoELinearForCausalLM).
+
+Guards the hybrid linear / full attention dispatcher: Ling-2.5/2.6
+has 32 layers with `layer_group_size=8`, so layers {7, 15, 23, 31}
+are full attention (MLA) and the rest are linear (Lightning seg_la).
+Runs on the 8-GPU H200 runner with TP=4.
+"""
+
+import unittest
+
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
+from sglang.test.server_fixtures.default_fixture import DefaultServerBase
+
+register_cuda_ci(est_time=600, stage="base-c", runner_config="8-gpu-h200")
+
+
+class TestLing26Flash(GSM8KMixin, DefaultServerBase):
+    model = "inclusionAI/Ling-2.6-flash"
+
+    # Native 128K context (no YaRN) — avoids the
+    # SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN env-var dance and keeps the
+    # smoke test focused on the dispatcher / hybrid-attention path.
+    other_args = [
+        "--tp-size",
+        "4",
+        "--trust-remote-code",
+        "--mamba-scheduler-strategy",
+        "extra_buffer",
+        "--mem-fraction-static",
+        "0.75",
+        "--max-running-requests",
+        "64",
+        "--max-mamba-cache-size",
+        "256",
+        # MTP path also exercises the dispatcher (draft + target verify),
+        # so keep it on to maximize coverage.
+        "--speculative-algorithm",
+        "NEXTN",
+        "--speculative-num-steps",
+        "3",
+        "--speculative-eagle-topk",
+        "1",
+        "--speculative-num-draft-tokens",
+        "4",
+    ]
+
+    # Observed 0.825 on H200 TP=4 + NEXTN MTP with default 200-question GSM8K
+    # (the model card's 0.96 is from full 1319-question runs of the 1T model).
+    gsm8k_accuracy_thres = 0.825
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=3)