From 76c9899da7f95893a5d86675b2dd6e1a9928d6e2 Mon Sep 17 00:00:00 2001
From: Alison Shao <54658187+alisonshao@users.noreply.github.com>
Date: Tue, 2 Jun 2026 16:24:49 -0700
Subject: [PATCH] Fix hybrid linear attention misrouting plain-RadixAttention
 linear layers to the full backend (Ring-2.5-1T) (#26623)

Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
---
 .../attention/hybrid_linear_attn_backend.py   | 21 +++++++------------
 .../sglang/srt/models/bailing_moe_linear.py   |  6 ------
 .../8-gpu-models/test_ling_2_6_flash.py       |  4 ++--
 3 files changed, 10 insertions(+), 21 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 71248dcf2..ac5a5e57a 100644
--- a/python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
+++ b/python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
@@ -782,22 +782,17 @@ class HybridLinearAttnBackend(AttentionBackend):
         self.req_to_token_pool = full_attn_backend.req_to_token_pool
 
     def _is_full_attn(
-        self, layer: Optional[RadixAttention], layer_id: Optional[int] = None
+        self,
+        layer: Optional[Union[RadixAttention, RadixLinearAttention]],
+        layer_id: Optional[int] = None,
     ) -> bool:
-        # Explicit linear-attention subclass → strong linear signal (KDA, GDN,
-        # Qwen3-Next, Qwen3.5 main linear layers).
+        # 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.
         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 1c4a5d872..c0c6be3ca 100644
--- a/python/sglang/srt/models/bailing_moe_linear.py
+++ b/python/sglang/srt/models/bailing_moe_linear.py
@@ -508,12 +508,6 @@ 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))
diff --git a/test/registered/8-gpu-models/test_ling_2_6_flash.py b/test/registered/8-gpu-models/test_ling_2_6_flash.py
index 4044a7619..856ebe0ed 100644
--- a/test/registered/8-gpu-models/test_ling_2_6_flash.py
+++ b/test/registered/8-gpu-models/test_ling_2_6_flash.py
@@ -3,7 +3,7 @@
 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.
+Runs nightly on the 8-GPU H200 runner with TP=4.
 """
 
 import unittest
@@ -12,7 +12,7 @@ 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")
+register_cuda_ci(est_time=600, suite="nightly-8-gpu-common", nightly=True)
 
 
 class TestLing26Flash(GSM8KMixin, DefaultServerBase):