[EPLB] Add contention-aware expert migration batching (#52641)

Signed-off-by: DOCCA0 <2947323341@qq.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
DOCCA0
2026-10-03 00:13:50 +00:00
committed by GitHub
co-authored by Robert Shaw
parent 8e443871f0
commit 7dfe3338d5
6 changed files with 588 additions and 8 deletions
+11
View File
@@ -106,6 +106,17 @@ class EPLBConfig:
on CUDA, falls back to "torch_gloo")
"""
enable_migration_batching: bool = False
"""Schedule expert migrations in batches where each rank communicates with
at most one peer. This reduces per-rank network contention at the cost of
additional sequential communication steps. This option only applies to
async EPLB and is disabled by default."""
@property
def migration_batching_enabled(self) -> bool:
"""Whether contention-aware batching is active for this configuration."""
return self.use_async and self.enable_migration_batching
@model_validator(mode="after")
def _validate_eplb_config(self) -> Self:
if self.use_async and self.policy != "default":