[CPU] Enable shared-memory based pipeline parallel for CPU backend (#21289)

Signed-off-by: jiang1.li <jiang1.li@intel.com>
This commit is contained in:
Li, Jiang
2025-07-21 09:07:08 -07:00
committed by GitHub
parent 6dda13c86b
commit a15a50fc17
8 changed files with 165 additions and 59 deletions
+12
View File
@@ -272,6 +272,9 @@ class GroupCoordinator:
self.use_custom_op_call = (current_platform.is_cuda_alike()
or current_platform.is_tpu())
self.use_cpu_custom_send_recv = (current_platform.is_cpu() and hasattr(
torch.ops._C, "init_shm_manager"))
@property
def first_rank(self):
"""Return the global rank of the first process in the group"""
@@ -663,6 +666,11 @@ class GroupCoordinator:
dst = (self.rank_in_group + 1) % self.world_size
assert dst < self.world_size, f"Invalid dst rank ({dst})"
if self.use_cpu_custom_send_recv:
self.device_communicator.send_tensor_dict( # type: ignore
tensor_dict, dst)
return None
metadata_list: list[tuple[Any, Any]] = []
assert isinstance(
tensor_dict,
@@ -718,6 +726,10 @@ class GroupCoordinator:
src = (self.rank_in_group - 1) % self.world_size
assert src < self.world_size, f"Invalid src rank ({src})"
if self.use_cpu_custom_send_recv:
return self.device_communicator.recv_tensor_dict( # type: ignore
src)
recv_metadata_list = self.recv_object(src=src)
tensor_dict: dict[str, Any] = {}
for key, value in recv_metadata_list: