minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Fix dpsk-r1-fp4 launching crash (#12063)

Browse Source
This commit is contained in:
Qiaolin Yu
2025-10-24 17:04:50 +08:00
committed by GitHub
parent b9fb74f3bc
commit 71d41212e4
2 changed files with 11 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/quantization/mxfp4.py
+5 -7
View File
@@ -587,13 +587,11 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
): ):
self.moe_runner_config = moe_runner_config self.moe_runner_config = moe_runner_config
backend = get_moe_runner_backend() backend = (
if backend.is_auto(): MoeRunnerBackend.TRITON_KERNELS
backend = ( if self.use_triton_kernels
MoeRunnerBackend.TRITON_KERNELS else MoeRunnerBackend.TRITON
if self.use_triton_kernels )
else MoeRunnerBackend.TRITON
)
self.runner = MoeRunner(backend, moe_runner_config) self.runner = MoeRunner(backend, moe_runner_config)
def apply( def apply(
python/sglang/srt/layers/quantization/unquant.py
+6 -13
View File
@@ -8,12 +8,7 @@ from torch.nn.parameter import Parameter
from sglang.srt.custom_op import CustomOp from sglang.srt.custom_op import CustomOp
from sglang.srt.layers.amx_utils import _amx_process_weight_after_loading from sglang.srt.layers.amx_utils import _amx_process_weight_after_loading
from sglang.srt.layers.moe import ( from sglang.srt.layers.moe import MoeRunner, MoeRunnerBackend, MoeRunnerConfig
MoeRunner,
MoeRunnerBackend,
MoeRunnerConfig,
get_moe_runner_backend,
)
from sglang.srt.layers.moe.moe_runner.triton import TritonMoeQuantInfo from sglang.srt.layers.moe.moe_runner.triton import TritonMoeQuantInfo
from sglang.srt.layers.quantization.base_config import ( from sglang.srt.layers.quantization.base_config import (
FusedMoEMethodBase, FusedMoEMethodBase,
@@ -225,13 +220,11 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
): ):
self.moe_runner_config = moe_runner_config self.moe_runner_config = moe_runner_config
backend = get_moe_runner_backend() backend = (
if backend.is_auto(): MoeRunnerBackend.TRITON_KERNELS
backend = ( if self.use_triton_kernels
MoeRunnerBackend.TRITON_KERNELS else MoeRunnerBackend.TRITON
if self.use_triton_kernels )
else MoeRunnerBackend.TRITON
)
self.runner = MoeRunner(backend, moe_runner_config) self.runner = MoeRunner(backend, moe_runner_config)
def apply( def apply(