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

Support redundant experts in expert parallel (#6461)

Browse Source
This commit is contained in:
fzyzcjy
2025-05-21 17:05:53 +08:00
committed by GitHub
parent a071dc4084
commit ccfe5c009d
5 changed files with 18 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
python/sglang/srt/server_args.py
View File

@@ -170,6 +170,7 @@ class ServerArgs:
enable_ep_moe: bool = False
enable_deepep_moe: bool = False
deepep_mode: Optional[Literal["auto", "normal", "low_latency"]] = "auto"
ep_num_redundant_experts: int = 0
ep_dispatch_algorithm: Optional[Literal["static", "dynamic"]] = None
init_expert_location: str = "trivial"
expert_distribution_recorder_mode: Optional[
@@ -1273,6 +1274,12 @@ class ServerArgs:
default="auto",
help="Select the mode when enable DeepEP MoE, could be `normal`, `low_latency` or `auto`. Default is `auto`, which means `low_latency` for decode batch and `normal` for prefill batch.",
)
parser.add_argument(
"--ep-num-redundant-experts",
type=int,
default=ServerArgs.ep_num_redundant_experts,
help="Allocate this number of redundant experts in expert parallel.",
)
parser.add_argument(
"--ep-dispatch-algorithm",
type=str,