[Elastic EP] Reduce eager-mode reconfiguration downtime (#51885)

Signed-off-by: Itay Alroy <ialroy@nvidia.com>
This commit is contained in:
Itay Alroy
2026-08-19 02:13:06 +00:00
committed by GitHub
parent d36cc4254e
commit a2257f95b7
12 changed files with 194 additions and 303 deletions
+5 -15
View File
@@ -3517,17 +3517,11 @@ class GPUModelRunner(
def setup_eplb_from_mapping(
self,
expanded_physical_to_logical: torch.Tensor,
old_num_physical_experts: int,
) -> None:
assert self._moe_model is not None
self.eplb_state = EplbState.from_mapping(
model=self._moe_model,
model_config=self.model_config,
device=self.device,
parallel_config=self.parallel_config,
expanded_physical_to_logical=expanded_physical_to_logical,
num_valid_physical_experts=old_num_physical_experts,
assert self.eplb_state is not None
self.eplb_state.update_mapping(
self.model_config,
expanded_physical_to_logical,
)
def _pool(
@@ -5488,11 +5482,7 @@ class GPUModelRunner(
# MixtureOfExperts themselves.
self._moe_model = get_mixture_of_experts_model(self.model)
if (
self.parallel_config.enable_eplb
and not load_dummy_weights
and self._moe_model is not None
):
if self._moe_model is not None and self.parallel_config.enable_eplb:
logger.info_once(
"EPLB is enabled for MoE part of model %s.",
self.model_config.model,