本文演示了tensor并行的原理。如何将二个mlp切分到多张GPU上分别计算自己的分块,最后做一次reduce。
1.为了避免中间数据产生集合通信,A矩阵只能按列切分,算出全部batch*seqlen的部分feature
2.因上面的步骤每张GPU只有部分feature,B矩阵可按行切分,与之进行矩阵乘,生成部分和
3.最后累加每张GPU上的部分和,即最终的结果
以下demo,先运行非分块的版本,然后模拟分块,最后是分布式的实现

一.原理

二.实现代码

python"># torch_tp_demo.py
import os
import torch
from torch import nn
import torch.nn.functional as F 
import numpy as np
import torch.distributed as dist
from torch.distributed import ReduceOp
  
import time
import argparse

parser = argparse.ArgumentParser(description="")
parser.add_argument('--hidden_size', default=512, type=int, help='')
parser.add_argument('--ffn_size', default=1024, type=int, help='')
parser.add_argument('--seq_len', default=512, type=int, help='')
parser.add_argument('--batch_size', default=8, type=int, help='')
parser.add_argument('--world_size', default=4, type=int, help='')
parser.add_argument('--device', default="cuda", type=str, help='')

class FeedForward(nn.Module): 

    def __init__(self,hidden_size,ffn_size): 
        super(FeedForward, self).__init__() 
        self.fc1 = nn.Linear(hidden_size, ffn_size,bias=False)
        self.fc2 = nn.Linear(ffn_size, hidden_size,bias=False)

    def forward(self, input): 
        return self.fc2(self.fc1(input))

class FeedForwardTp(nn.Module):

    def __init__(self,hidden_size,ffn_size,tp_size,rank): 
        super(FeedForwardTp, self).__init__() 
        self.fc1 = nn.Linear(hidden_size, ffn_size//tp_size,bias=False)
        self.fc2 = nn.Linear(ffn_size//tp_size, hidden_size,bias=False)
      
        self.fc1.weight.data=torch.from_numpy(np.fromfile(f"fc1_{rank}.bin",dtype=np.float32)).reshape(self.fc1.weight.data.shape)
        self.fc2.weight.data=torch.from_numpy(np.fromfile(f"fc2_{rank}.bin",dtype=np.float32)).reshape(self.fc2.weight.data.shape)

    def forward(self, input): 
        return self.fc2(self.fc1(input))


args = parser.parse_args()
hidden_size = args.hidden_size
ffn_size = args.ffn_size
seq_len = args.seq_len
batch_size = args.batch_size
world_size = args.world_size
device = args.device

def native_mode():
  print(args)
  torch.random.manual_seed(1)
  model = FeedForward(hidden_size,ffn_size)
  model.eval()
  input = torch.rand((batch_size, seq_len, hidden_size),dtype=torch.float32).half().to(device)

  for idx,chunk in enumerate(torch.split(model.fc1.weight, ffn_size//world_size, dim=0)):
      chunk.data.numpy().tofile(f"fc1_{idx}.bin")
  
  for idx,chunk in enumerate(torch.split(model.fc2.weight, ffn_size//world_size, dim=1)):
      chunk.data.numpy().tofile(f"fc2_{idx}.bin")
  
  model=model.half().to(device)
  

  usetime=[]
  for i in range(32):
    t0=time.time()    
    out = model(input)
    torch.cuda.synchronize()
    t1=time.time()
    if i>3:
      usetime.append(t1-t0)
  
  print("[INFO] native: shape:{},sum:{:.5f},mean:{:.5f},min:{:.5f},max:{:.5f}".format(out.shape,out.sum().item(),np.mean(usetime),np.min(usetime),np.max(usetime)))
  

  result=[]
  for rank in range(world_size):
      model = FeedForwardTp(hidden_size,ffn_size,world_size,rank).half().to(device)
      model.eval()
      out=model(input)
      torch.cuda.synchronize()
      result.append(out)
  
  sum_all=result[0]
  for t in result[1:]:
      sum_all=sum_all+t
  
  print("[INFO] tp_simulate: shape:{},sum:{:.5f}".format(sum_all.shape,sum_all.sum().item()))

def tp_mode():
  torch.random.manual_seed(1)
  dist.init_process_group(backend='nccl')
    
  world_size = torch.distributed.get_world_size()
  rank=rank = torch.distributed.get_rank()
  local_rank=int(os.environ['LOCAL_RANK'])
  
  torch.cuda.set_device(local_rank)
  device = torch.device("cuda",local_rank)
  
  input = torch.rand((batch_size, seq_len, hidden_size),dtype=torch.float32).half().to(device)  
  model = FeedForwardTp(hidden_size,ffn_size,world_size,rank).half().to(device)
  model.eval()
  if rank==0:
    print(args)
    
  usetime=[]
  for i in range(32):        
    dist.barrier()
    t0=time.time()
    out=model(input)
    #dist.reduce(out,0, op=ReduceOp.SUM) 
    dist.all_reduce(out,op=ReduceOp.SUM)
    torch.cuda.synchronize()
    if rank==0:
      t1=time.time()
      if i>3:
        usetime.append(t1-t0)
  
  if rank==0:
    print("[INFO] tp: shape:{},sum:{:.5f},mean:{:.5f},min:{:.5f},max:{:.5f}".format(out.shape,out.sum().item(),np.mean(usetime),np.min(usetime),np.max(usetime)))


if __name__ == "__main__":
  num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
  is_distributed = num_gpus > 1
  if is_distributed:
    tp_mode()
  else:
    native_mode()

运行命令:

python3 torch_tp_demo.py --hidden_size 512 \
			--ffn_size 4096 --seq_len 512 \
			--batch_size 8 --world_size 4 --device "cuda"
torchrun -m --nnodes=1 --nproc_per_node=4 \
			torch_tp_demo --hidden_size 512 \
			--ffn_size 4096 --seq_len 512 \
			--batch_size 8 --world_size 4