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

Tiny support customize DeepEP max dispatch tokens per rank (#6934)

Browse Source
This commit is contained in:
fzyzcjy
2025-06-07 08:18:35 +08:00
committed by GitHub
parent e6b7053b60
commit 61ce91ed28
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/layers/moe/ep_moe/token_dispatcher.py
View File

@@ -6,7 +6,7 @@ from sglang.srt.managers.expert_distribution import (
get_global_expert_distribution_recorder,
)
from sglang.srt.managers.schedule_batch import global_server_args_dict
from sglang.srt.utils import DeepEPMode, load_json_config
from sglang.srt.utils import DeepEPMode, get_int_env_var, load_json_config
try:
from deep_ep import Buffer, Config
@@ -188,7 +188,9 @@ class _DeepEPDispatcherImplBase:
self.deepep_mode = deepep_mode
self.params_bytes = 2
self.num_max_dispatch_tokens_per_rank = 128
self.num_max_dispatch_tokens_per_rank = get_int_env_var(
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK", 128
)
self.handle = None