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

[NPU] ascend fuseep use moe ep group (#32040)

Browse Source
This commit is contained in:
heziiop
2026-07-23 14:07:14 +08:00
committed by GitHub
parent 09071be105
commit 235a488c87
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/hardware_backend/npu/moe/fuseep.py
+2 -2
View File
@@ -12,7 +12,7 @@ from typing import TYPE_CHECKING
import torch import torch
from sglang.srt.distributed import get_tp_group from sglang.srt.distributed import get_moe_ep_group
from sglang.srt.environ import envs from sglang.srt.environ import envs
from sglang.srt.hardware_backend.npu.utils import npu_format_cast from sglang.srt.hardware_backend.npu.utils import npu_format_cast
from sglang.srt.layers.moe.token_dispatcher.deepep import DeepEPBuffer from sglang.srt.layers.moe.token_dispatcher.deepep import DeepEPBuffer
@@ -30,7 +30,7 @@ _PARAMS_BYTES = 2 # bf16 — Ascend's Dispatch & Combine does not support fp16
def _get_fuseep_buffer(layer: FusedMoE): def _get_fuseep_buffer(layer: FusedMoE):
DeepEPBuffer.set_dispatch_mode_as_low_latency() DeepEPBuffer.set_dispatch_mode_as_low_latency()
return DeepEPBuffer.get_deepep_buffer( return DeepEPBuffer.get_deepep_buffer(
get_tp_group().device_group, get_moe_ep_group().device_group,
layer.hidden_size, layer.hidden_size,
_PARAMS_BYTES, _PARAMS_BYTES,
DeepEPMode.LOW_LATENCY, DeepEPMode.LOW_LATENCY,