mirror of
https://github.com/microsoft/mscclpp.git
synced 2026-05-11 17:00:22 +00:00
bd68319e3e
This PR refactors the algorithm selection logic in MSCCL++ and introduces support for symmetric memory configuration through environment variables. 1. Algorithm Selection Refactoring Use separate class for algo selection. Could introduce more complex logic for algo selection based on message size, arch, if cuda graph is enabled and memory allocation method 2. Symmetric Memory Support Introduced symmetricMemory parameter in algorithm context key generation. Remove disableChannelCache env as is ambiguous 3. Add new args for build_default_algorithms Add flag_buffer, and flag_buffer_size args to build default algorithm. Then we could use unified flag buffer for different algorithms, avoid application hanging when switch algo for different message size. --------- Co-authored-by: chhwang <8018170+chhwang@users.noreply.github.com> Co-authored-by: Qinghua Zhou <qinghuazhou@microsoft.com> Co-authored-by: Caio Rocha <caiorocha@microsoft.com>
83 lines
2.7 KiB
Python
83 lines
2.7 KiB
Python
# Copyright (c) Microsoft Corporation.
|
|
# Licensed under the MIT License.
|
|
|
|
# LD_PRELOAD=<MSCCLPP_REPO>/build/lib/libmscclpp_nccl.so MSCCLPP_NCCL_SYMMETRIC_MEMORY=false torchrun --nnodes=1 --nproc_per_node=8 memory_report.py
|
|
import os, sys
|
|
import torch
|
|
import torch.distributed as dist
|
|
|
|
|
|
def memory_report(d) -> str:
|
|
"""
|
|
One-line CUDA memory report for the current device.
|
|
"""
|
|
if not torch.cuda.is_available():
|
|
return "MEMORY REPORT: CUDA not available"
|
|
torch.cuda.synchronize(d)
|
|
|
|
allocated = torch.cuda.memory_allocated(d)
|
|
reserved = torch.cuda.memory_reserved(d)
|
|
max_alloc = torch.cuda.max_memory_allocated(d)
|
|
max_resv = torch.cuda.max_memory_reserved(d)
|
|
|
|
free_b, total_b = torch.cuda.mem_get_info(d) # (free, total) in bytes
|
|
used_b = total_b - free_b
|
|
|
|
to_gib = lambda b: f"{b / (1024**3):.2f} GiB"
|
|
return (
|
|
"MEMORY REPORT: "
|
|
f"torch allocated: {to_gib(allocated)} | "
|
|
f"torch reserved: {to_gib(reserved)} | "
|
|
f"max torch allocated: {to_gib(max_alloc)} | "
|
|
f"max torch reserved: {to_gib(max_resv)} | "
|
|
f"total memory used: {to_gib(used_b)} | "
|
|
f"total memory: {to_gib(total_b)}"
|
|
)
|
|
|
|
|
|
def main():
|
|
# torchrun provides these envs
|
|
rank = int(os.environ["RANK"])
|
|
world_size = int(os.environ["WORLD_SIZE"])
|
|
local_rank = int(os.environ["LOCAL_RANK"])
|
|
|
|
nelems = 1024 * 1024 * 32 # 32M elements
|
|
torch.cuda.set_device(local_rank)
|
|
backend = "nccl"
|
|
|
|
# init default PG
|
|
dist.init_process_group(backend=backend, init_method="env://")
|
|
if rank == 0:
|
|
print(
|
|
f"[world_size={world_size}] torch={torch.__version__}, cuda={torch.version.cuda}, backend={backend}",
|
|
flush=True,
|
|
)
|
|
dist.barrier()
|
|
|
|
# make a subgroup over all ranks (you can change to a subset to test)
|
|
group_ranks = list(range(world_size))
|
|
if rank == 0:
|
|
print(f"Creating new_group with ranks={group_ranks}", flush=True)
|
|
grp0 = dist.new_group(ranks=group_ranks, backend=backend)
|
|
x = torch.ones(nelems, device=local_rank, dtype=torch.float32) * (rank + 1)
|
|
dist.all_reduce(x, op=dist.ReduceOp.SUM, group=grp0)
|
|
|
|
grp1 = dist.new_group(ranks=list(range(world_size)), backend=backend)
|
|
x = torch.ones(nelems, device=local_rank, dtype=torch.float32) * (rank + 1)
|
|
dist.all_reduce(x, op=dist.ReduceOp.SUM, group=grp1)
|
|
|
|
dist.barrier()
|
|
|
|
print(memory_report(local_rank))
|
|
dist.destroy_process_group(grp0)
|
|
dist.destroy_process_group(grp1)
|
|
dist.destroy_process_group()
|
|
|
|
|
|
if __name__ == "__main__":
|
|
try:
|
|
main()
|
|
except Exception as e:
|
|
print(f"[rank {os.getenv('RANK','?')}] EXCEPTION: {e}", file=sys.stderr, flush=True)
|
|
raise
|