Skip to content

Use symmetric memory in matmul+allreduce reference implementation #5837

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wujingyue wants to merge 2 commits into
base: main
Choose a base branch
from
Draft

Use symmetric memory in matmul+allreduce reference implementation #5837

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
cb4d494
Reference model uses num_chunks=4 and num_streams=2
wujingyue Jan 17, 2026
b286468
Use symmetric memory
wujingyue Jan 17, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 11 additions & 2 deletions tests/python/multidevice/conftest.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,7 @@

import torch
import torch.distributed as dist
import torch.distributed._symmetric_memory as symm_mem

import nvfuser_direct as nvfuser

Expand Down Expand Up @@ -133,13 +134,21 @@ def setup_default_process_group():

torch.cuda.set_device(local_rank)

# The default port as used by https://github.com/pytorch/pytorch/blob/45a8b5682eb69d865cbf68c7f2f689b56b4efd53/torch/csrc/distributed/c10d/TCPStore.hpp#L51.
opts = dist.ProcessGroupNCCL.Options()
opts.config.cta_policy = dist.ProcessGroupNCCL.NCCL_CTA_POLICY_ZERO
dist.init_process_group(
backend="nccl",
pg_options=opts,
# The default port as used by https://github.com/pytorch/pytorch/blob/45a8b5682eb69d865cbf68c7f2f689b56b4efd53/torch/csrc/distributed/c10d/TCPStore.hpp#L51.
init_method="tcp://localhost:29500",
world_size=world_size,
rank=rank,
device_id=torch.device(f"cuda:{local_rank}"),
device_id=local_rank,
)

symm_mem.set_backend("NCCL")
symm_mem.enable_symm_mem_for_group(dist.group.WORLD.group_name)

yield

dist.destroy_process_group()
31 changes: 21 additions & 10 deletions tests/python/multidevice/test_overlap.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,7 @@

import torch
import torch.distributed as dist
import torch.distributed._symmetric_memory as symm_mem
from torch.distributed.tensor import distribute_tensor, Shard

import nvfuser_direct as nvfuser
Expand Down Expand Up @@ -168,22 +169,17 @@ def get(self, sid: int) -> torch.cuda.Stream:
def row_parallel_linear_forward_reference(
inp_shard: torch.Tensor,
weight_shard: torch.Tensor,
out: torch.Tensor,
num_chunks: int,
stream_pool: StreamPool,
) -> torch.Tensor:
out = torch.empty(
inp_shard.size(0),
weight_shard.size(0),
device="cuda",
dtype=inp_shard.dtype,
)
inp_chunks = inp_shard.chunk(num_chunks)
out_chunks = out.chunk(num_chunks)

main_stream = torch.cuda.current_stream()
worker_streams = []
for i, (inp_chunk, out_chunk) in enumerate(zip(inp_chunks, out_chunks)):
worker_stream = stream_pool.get(i)
worker_stream = stream_pool.get(i % 2)
worker_streams.append(worker_stream)
worker_stream.wait_stream(main_stream)
with torch.cuda.stream(worker_stream):
Expand Down Expand Up @@ -222,7 +218,15 @@ def test_row_parallel_linear_forward_reference(setup_default_process_group):
weight_ref, mesh, placements=[Shard(-1)]
).to_local()
stream_pool = StreamPool()
out = row_parallel_linear_forward_reference(inp_shard, weight_shard, s, stream_pool)

out = symm_mem.empty(
inp_shard.size(0),
weight_shard.size(0),
device="cuda",
dtype=inp_shard.dtype,
)
symm_mem.rendezvous(out, group=dist.group.WORLD)
row_parallel_linear_forward_reference(inp_shard, weight_shard, out, s, stream_pool)

torch.testing.assert_close(out.cpu(), out_ref)

Expand All @@ -232,7 +236,7 @@ def test_row_parallel_linear_forward_reference(setup_default_process_group):
def test_row_parallel_linear_forward_reference_benchmark(
setup_default_process_group, benchmark
):
h, s, t = 8192, 2, 8192
h, s, t = 8192, 4, 8192
d = dist.get_world_size()
if (h * 4) % d != 0:
pytest.skip(
Expand All @@ -251,9 +255,16 @@ def test_row_parallel_linear_forward_reference_benchmark(
).to_local()

stream_pool = StreamPool()
out = symm_mem.empty(
inp_shard.size(0),
weight_shard.size(0),
device="cuda",
dtype=inp_shard.dtype,
)
symm_mem.rendezvous(out, group=dist.group.WORLD)
warmup_fn, benchmark_fn = get_benchmark_fns(
lambda: row_parallel_linear_forward_reference(
inp_shard, weight_shard, s, stream_pool
inp_shard, weight_shard, out, s, stream_pool
)
)
warmup_fn()
Expand Down