[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
-2
View File
@@ -144,7 +144,6 @@ class EPLBController:
model: nn.Module,
model_config: Any,
expanded_physical_to_logical: torch.Tensor,
old_num_physical_experts: int,
) -> None:
moe_model = get_mixture_of_experts_model(model)
assert moe_model is not None
@@ -155,6 +154,5 @@ class EPLBController:
device=self.device,
parallel_config=self.parallel_config,
expanded_physical_to_logical=expanded_physical_to_logical,
num_valid_physical_experts=old_num_physical_experts,
)
self._has_registered_models = True
-2
View File
@@ -1962,13 +1962,11 @@ class GPUModelRunner(LoRAModelRunnerMixin):
def setup_eplb_from_mapping(
self,
expanded_physical_to_logical: torch.Tensor,
old_num_physical_experts: int,
) -> None:
self.eplb.setup_from_mapping(
self.model,
self.model_config,
expanded_physical_to_logical,
old_num_physical_experts,
)
########### EPLB methods end ###########