[MoE Refactor] Integrate Naive Prepare Finalize into MK (#32567)

Signed-off-by: Robert Shaw <robshaw@redhat.com>
Signed-off-by: Amir Klein <203507526+amirkl94@users.noreply.github.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: amirkl94 <203507526+amirkl94@users.noreply.github.com>
This commit is contained in:
Robert Shaw
2026-01-27 01:28:02 +00:00
committed by GitHub
co-authored by Robert Shaw amirkl94
parent 6d86fde09c
commit 5a93b9162b
46 changed files with 1018 additions and 876 deletions
+24 -2
View File
@@ -1000,7 +1000,7 @@ class GroupCoordinator:
if self.device_communicator is not None:
self.device_communicator.prepare_communication_buffer_for_model(model)
def dispatch(
def dispatch_router_logits(
self,
hidden_states: torch.Tensor,
router_logits: torch.Tensor,
@@ -1011,7 +1011,7 @@ class GroupCoordinator:
| tuple[torch.Tensor, torch.Tensor, list[torch.Tensor]]
):
if self.device_communicator is not None:
return self.device_communicator.dispatch( # type: ignore[call-arg]
return self.device_communicator.dispatch_router_logits(
hidden_states,
router_logits,
is_sequence_parallel,
@@ -1020,6 +1020,28 @@ class GroupCoordinator:
else:
return hidden_states, router_logits
def dispatch(
self,
hidden_states: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
is_sequence_parallel: bool = False,
extra_tensors: list[torch.Tensor] | None = None,
) -> (
tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[torch.Tensor]]
| tuple[torch.Tensor, torch.Tensor, torch.Tensor]
):
if self.device_communicator is not None:
return self.device_communicator.dispatch(
hidden_states,
topk_weights,
topk_ids,
is_sequence_parallel,
extra_tensors,
)
else:
return hidden_states, topk_weights, topk_ids
def combine(
self, hidden_states, is_sequence_parallel: bool = False
) -> torch.Tensor: