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

Fix hopper launch gpt-oss model illegal memory (#8908)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2025-08-08 01:02:40 +08:00
committed by GitHub
parent 500b15c960
commit 3ae33fcd0a
1 changed files with 13 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
python/sglang/srt/layers/quantization/mxfp4.py
View File

@@ -16,6 +16,7 @@ from sglang.srt.layers.quantization.base_config import (
QuantizeMethodBase,
)
from sglang.srt.layers.quantization.utils import is_layer_skipped
from sglang.srt.layers.utils import is_sm100_supported
from sglang.srt.managers.schedule_batch import global_server_args_dict
from sglang.srt.utils import (
direct_register_custom_op,
@@ -28,6 +29,7 @@ from sglang.srt.utils import (
set_weight_attrs,
)
_is_sm100_supported = is_cuda() and is_sm100_supported()
has_triton_kernels = importlib.util.find_spec("triton_kernels") is not None
@@ -244,13 +246,17 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
# pad the intermediate size to be a multiple of 2 * mxfp4_block
# for to hold non-uniform sharded tensor as well as swizzling
if self.use_flashinfer:
intermediate_size_per_partition_after_pad = round_up(intermediate_size, 256)
hidden_size = round_up(hidden_size, 256)
elif is_hip():
intermediate_size_per_partition_after_pad = round_up(intermediate_size, 128)
else:
intermediate_size_per_partition_after_pad = round_up(intermediate_size, 64)
intermediate_size_per_partition_after_pad = intermediate_size
if _is_sm100_supported:
if self.use_flashinfer:
intermediate_size_per_partition_after_pad = round_up(
intermediate_size, 256
)
hidden_size = round_up(hidden_size, 256)
else:
intermediate_size_per_partition_after_pad = round_up(
intermediate_size, 64
)
self.intermediate_size = intermediate_size_per_partition_after_pad