EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Deprecate global_server_args_dict (#11331)

Browse Source
This commit is contained in:
Liangsheng Yin
2025-10-13 01:20:47 +08:00
committed by GitHub
parent 2157d12ae8
commit 1083e7e3df
54 changed files with 240 additions and 321 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
python/sglang/srt/layers/quantization/mxfp4.py
View File

@@ -31,7 +31,7 @@ from sglang.srt.layers.quantization.base_config import (
QuantizeMethodBase,
)
from sglang.srt.layers.quantization.utils import is_layer_skipped
from sglang.srt.managers.schedule_batch import global_server_args_dict
from sglang.srt.server_args import get_global_server_args
from sglang.srt.utils import (
direct_register_custom_op,
is_cuda,
@@ -265,9 +265,9 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
self.use_triton_kernels = get_moe_runner_backend().is_triton_kernel()
self.with_bias = False
self.use_flashinfer = get_moe_runner_backend().is_flashinfer_mxfp4()
self.flashinfer_mxfp4_moe_precision = global_server_args_dict[
"flashinfer_mxfp4_moe_precision"
]
self.flashinfer_mxfp4_moe_precision = (
get_global_server_args().flashinfer_mxfp4_moe_precision
)
self.triton_kernel_moe_forward = None
self.triton_kernel_moe_with_bias_forward = None