Making large AI models cheaper, faster and more accessible
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 

83 lines
2.5 KiB

import pytest
import torch
import torch.distributed as dist
import colossalai
from colossalai.accelerator import get_accelerator
from colossalai.cluster import ProcessGroupMesh
from colossalai.pipeline.p2p import PipelineP2PCommunication, create_send_metadata
from colossalai.pipeline.stage_manager import PipelineStageManager
from colossalai.testing import rerun_if_address_is_in_use, spawn
WORLD_SIZE = 2
def check_p2p_communication():
pg_mesh = ProcessGroupMesh(WORLD_SIZE)
stage_manager = PipelineStageManager(pg_mesh, 0)
p2p = PipelineP2PCommunication(stage_manager, overlap_p2p=False)
rank = dist.get_rank()
tensor = torch.ones(1, device=get_accelerator().get_current_device())
data = [
"tensor",
tensor,
[tensor],
{"tensor": tensor},
]
if rank == 0:
for obj in data:
p2p.send_forward(obj)
for i in range(len(data)):
recv_obj, _ = p2p.send_forward_recv_backward(data[i], send_first=False)
assert recv_obj == data[-(i + 1)]
elif rank == 1:
for obj in data:
recv_obj, _ = p2p.recv_forward()
assert recv_obj == obj
for i in range(len(data)):
p2p.send_backward(data[-(i + 1)])
recv_obj, _ = p2p.recv_forward()
assert recv_obj == data[i]
if rank == 1:
for obj in data:
p2p.send_backward(obj)
for i in range(len(data)):
recv_obj, _ = p2p.send_backward_recv_forward(data[i], send_first=True)
assert recv_obj == data[-(i + 1)]
elif rank == 0:
for obj in data:
recv_obj, _ = p2p.recv_backward()
assert recv_obj == obj
for i in range(len(data)):
recv_obj, _ = p2p.send_forward_recv_backward(data[-(i + 1)], send_first=False)
assert recv_obj == data[i]
if rank == 0:
recv_obj, _ = p2p.send_forward_recv_backward(
tensor,
send_metadata=False,
metadata_recv=create_send_metadata(tensor),
)
assert recv_obj == tensor
elif rank == 1:
recv_obj, _ = p2p.recv_forward(metadata_recv=create_send_metadata(tensor))
assert recv_obj == tensor
p2p.send_backward(tensor, send_metadata=False)
def run_dist(rank, world_size, port):
colossalai.launch(rank=rank, world_size=world_size, port=port, host="localhost")
check_p2p_communication()
@pytest.mark.dist
@rerun_if_address_is_in_use()
def test_pipeline_p2p():
spawn(run_dist, WORLD_SIZE)
if __name__ == "__main__":
test_pipeline_p2p()