mirror of
https://github.com/microsoft/mscclpp.git
synced 2026-05-12 17:26:04 +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>
92 lines
2.6 KiB
Python
92 lines
2.6 KiB
Python
# Copyright (c) Microsoft Corporation.
|
|
# Licensed under the MIT License.
|
|
|
|
# run with:
|
|
# LD_PRELOAD=<MSCCLPP_REPO>/build/lib/libmscclpp_nccl.so MSCCLPP_NCCL_SYMMETRIC_MEMORY=false torchrun --nproc_per_node=8 ./allreduce_temp_buff.py
|
|
|
|
import os
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.distributed as dist
|
|
|
|
|
|
def init_dist():
|
|
rank = int(os.environ["RANK"])
|
|
world = int(os.environ["WORLD_SIZE"])
|
|
local = int(os.environ["LOCAL_RANK"])
|
|
torch.cuda.set_device(local)
|
|
dist.init_process_group("nccl")
|
|
return rank, world, local
|
|
|
|
|
|
class SimpleModel(nn.Module):
|
|
def __init__(self, DIN, DH, DOUT):
|
|
super().__init__()
|
|
self.layer1 = nn.Linear(DIN, DH, bias=False)
|
|
self.layer2 = nn.Linear(DH, DOUT, bias=False)
|
|
self.rank = int(os.environ["RANK"])
|
|
self.eval()
|
|
|
|
@torch.no_grad()
|
|
def forward(self, x_bf16: torch.Tensor, out_bf16: torch.Tensor):
|
|
"""
|
|
x_bf16: [B, DIN] (bf16) input
|
|
out_bf16: [B, DOUT] (bf16) output buffer
|
|
Returns:
|
|
out_bf16: [B, DOUT] (bf16) output
|
|
"""
|
|
out = self.layer1(x_bf16)
|
|
temp = torch.empty_like(out, dtype=torch.bfloat16)
|
|
temp.copy_(out)
|
|
dist.all_reduce(temp, op=dist.ReduceOp.SUM)
|
|
temp2 = temp
|
|
if self.rank == 0:
|
|
# If we are on rank 0, we can use a different temp buffer, make sure msccl++ can handle buffer address changes
|
|
temp2 = torch.empty_like(temp, dtype=torch.bfloat16)
|
|
temp2.copy_(temp)
|
|
dist.all_reduce(temp2, op=dist.ReduceOp.SUM)
|
|
out = self.layer2(temp2)
|
|
out_bf16.copy_(out)
|
|
return out
|
|
|
|
|
|
def main():
|
|
rank, _, local = init_dist()
|
|
device = torch.device("cuda", local)
|
|
torch.set_grad_enabled(False)
|
|
|
|
# message size B * DH * dtype_size = 32MB
|
|
B, DIN, DH, DOUT = 2048, 1024, 8192, 8
|
|
dtype = torch.bfloat16
|
|
|
|
# Warm up comms
|
|
dist.all_reduce(torch.ones(1, device=device).to(dtype))
|
|
|
|
# Build model
|
|
model = SimpleModel(DIN, DH, DOUT).to(device).to(dtype)
|
|
|
|
# Static I/O buffers for capture (stable addresses)
|
|
x_bf16 = torch.empty(B, DIN, dtype=dtype, device=device)
|
|
out_bf16 = torch.empty(B, DOUT, dtype=dtype, device=device)
|
|
|
|
# Eager warmup
|
|
x_bf16.normal_()
|
|
_ = model(x_bf16, out_bf16)
|
|
torch.cuda.synchronize()
|
|
|
|
g = torch.cuda.CUDAGraph()
|
|
with torch.cuda.graph(g):
|
|
model(x_bf16, out_bf16)
|
|
|
|
for step in range(5):
|
|
x_bf16.normal_()
|
|
g.replay()
|
|
if rank == 0:
|
|
print(f"[step {step}] out_mean={out_bf16.float().mean().item():.6f}")
|
|
|
|
dist.destroy_process_group()
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|