[CPU][BugFix] Fix inter-node pipeline parallel (#40150)

Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
This commit is contained in:
Fadi Arafeh
2026-04-20 17:21:12 +08:00
committed by GitHub
parent 58631d7c3f
commit 2aab9acf48
2 changed files with 17 additions and 2 deletions
@@ -45,6 +45,9 @@ class CpuCommunicator(DeviceCommunicatorBase):
unique_name,
)
# send/recv tensor_dict is only supported through the SHM communicator backend
self.supports_tensor_dict = isinstance(self.dist_module, _CPUSHMDistributed)
if self.use_all2all:
if self.all2all_backend != "naive": # type: ignore[has-type]
logger.warning(
@@ -143,12 +146,22 @@ class CpuCommunicator(DeviceCommunicatorBase):
tensor_dict: dict[str, torch.Tensor | Any],
dst: int,
) -> None:
if not self.supports_tensor_dict:
raise NotImplementedError(
"CpuCommunicator does not support tensor dict fastpath with "
"torch.distributed backend."
)
return self.dist_module.send_tensor_dict(tensor_dict, dst)
def recv_tensor_dict(
self,
src: int,
) -> dict[str, torch.Tensor | Any]:
if not self.supports_tensor_dict:
raise NotImplementedError(
"CpuCommunicator does not support tensor dict fastpath with "
"torch.distributed backend."
)
return self.dist_module.recv_tensor_dict(src)
def dispatch_router_logits(