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

[6/N] MoE Refactor: Cleanup MoE-related configs (#8849)

Browse Source
This commit is contained in:
Cheng Wan
2025-08-14 21:14:53 -07:00
committed by GitHub
parent 584e1ab2d0
commit 295895120d
69 changed files with 956 additions and 1037 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
python/sglang/srt/eplb/expert_distribution.py
View File

@@ -25,7 +25,6 @@ import torch
import torch.distributed
from sglang.srt.eplb.expert_location import ExpertLocationMetadata
from sglang.srt.managers.schedule_batch import global_server_args_dict
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
from sglang.srt.server_args import ServerArgs
from sglang.srt.utils import Withable, get_bool_env_var
@@ -288,14 +287,14 @@ class _SinglePassGatherer(ABC):
)
if server_args.expert_distribution_recorder_mode == "stat_approx":
if server_args.moe_a2a_backend is not None and (
if server_args.moe_a2a_backend != "none" and (
server_args.deepep_mode == "normal"
):
return _DeepepNormalSinglePassGatherer(expert_location_metadata, rank)
else:
raise NotImplementedError
if server_args.moe_a2a_backend is not None:
if server_args.moe_a2a_backend != "none":
if server_args.deepep_mode == "normal":
return _SelectExpertsSinglePassGatherer(expert_location_metadata, rank)
elif server_args.deepep_mode == "low_latency":