From 71d41212e4111ed335e5717d8151514d9c192b74 Mon Sep 17 00:00:00 2001
From: Qiaolin Yu <liin1211@outlook.com>
Date: Fri, 24 Oct 2025 02:04:50 -0700
Subject: [PATCH] Fix dpsk-r1-fp4 launching crash (#12063)

---
 .../sglang/srt/layers/quantization/mxfp4.py   | 12 +++++-------
 .../sglang/srt/layers/quantization/unquant.py | 19 ++++++-------------
 2 files changed, 11 insertions(+), 20 deletions(-)

diff --git a/python/sglang/srt/layers/quantization/mxfp4.py b/python/sglang/srt/layers/quantization/mxfp4.py
index d0db9a947..77f5f3f35 100644
--- a/python/sglang/srt/layers/quantization/mxfp4.py
+++ b/python/sglang/srt/layers/quantization/mxfp4.py
@@ -587,13 +587,11 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
         self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
     ):
         self.moe_runner_config = moe_runner_config
-        backend = get_moe_runner_backend()
-        if backend.is_auto():
-            backend = (
-                MoeRunnerBackend.TRITON_KERNELS
-                if self.use_triton_kernels
-                else MoeRunnerBackend.TRITON
-            )
+        backend = (
+            MoeRunnerBackend.TRITON_KERNELS
+            if self.use_triton_kernels
+            else MoeRunnerBackend.TRITON
+        )
         self.runner = MoeRunner(backend, moe_runner_config)
 
     def apply(
diff --git a/python/sglang/srt/layers/quantization/unquant.py b/python/sglang/srt/layers/quantization/unquant.py
index eb15dc442..55c5b8e3f 100644
--- a/python/sglang/srt/layers/quantization/unquant.py
+++ b/python/sglang/srt/layers/quantization/unquant.py
@@ -8,12 +8,7 @@ from torch.nn.parameter import Parameter
 
 from sglang.srt.custom_op import CustomOp
 from sglang.srt.layers.amx_utils import _amx_process_weight_after_loading
-from sglang.srt.layers.moe import (
-    MoeRunner,
-    MoeRunnerBackend,
-    MoeRunnerConfig,
-    get_moe_runner_backend,
-)
+from sglang.srt.layers.moe import MoeRunner, MoeRunnerBackend, MoeRunnerConfig
 from sglang.srt.layers.moe.moe_runner.triton import TritonMoeQuantInfo
 from sglang.srt.layers.quantization.base_config import (
     FusedMoEMethodBase,
@@ -225,13 +220,11 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
     ):
         self.moe_runner_config = moe_runner_config
-        backend = get_moe_runner_backend()
-        if backend.is_auto():
-            backend = (
-                MoeRunnerBackend.TRITON_KERNELS
-                if self.use_triton_kernels
-                else MoeRunnerBackend.TRITON
-            )
+        backend = (
+            MoeRunnerBackend.TRITON_KERNELS
+            if self.use_triton_kernels
+            else MoeRunnerBackend.TRITON
+        )
         self.runner = MoeRunner(backend, moe_runner_config)
 
     def apply(