feature(eplb): add min-rebalancing-utilization-threshold for eplb (#8345)

Co-authored-by: yizhang2077 <1109276519@qq.com>
This commit is contained in:
hzh0425
2025-08-30 11:24:29 +08:00
committed by GitHub
parent 591e6c5983
commit c2a26e725c
3 changed files with 62 additions and 4 deletions

View File

@@ -274,6 +274,7 @@ class ServerArgs:
eplb_algorithm: str = "auto"
eplb_rebalance_num_iterations: int = 1000
eplb_rebalance_layers_per_chunk: Optional[int] = None
eplb_min_rebalancing_utilization_threshold: float = 1.0
expert_distribution_recorder_mode: Optional[
Literal["stat", "stat_approx", "per_pass", "per_token"]
] = None
@@ -1595,6 +1596,12 @@ class ServerArgs:
default=ServerArgs.eplb_rebalance_layers_per_chunk,
help="Number of layers to rebalance per forward pass.",
)
parser.add_argument(
"--eplb-min-rebalancing-utilization-threshold",
type=float,
default=ServerArgs.eplb_min_rebalancing_utilization_threshold,
help="Minimum threshold for GPU average utilization to trigger EPLB rebalancing. Must be in the range [0.0, 1.0].",
)
parser.add_argument(
"--expert-distribution-recorder-mode",
type=str,