squall/picotron
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

leave out CP integration at the very end

Browse Source
This commit is contained in:
ferdinand.mom 2024-10-18 14:59:39 +00:00
parent d0d6d8994f
commit 83ddda2ce8
1 changed files with 7 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
train.py
View File

@ -8,7 +8,6 @@ CUDA_DEVICE_MAX_CONNECTIONS=1 torchrun --nproc_per_node=4 --nnodes=1 --
#VERBOSE=0 torchrun --nproc_per_node 4 --master_addr localhost --master_port 25500 train.py --pp_size 2 --dp_size 2 #VERBOSE=0 torchrun --nproc_per_node 4 --master_addr localhost --master_port 25500 train.py --pp_size 2 --dp_size 2
""" """
import multiprocessing
import os import os
import torch.nn.functional as F import torch.nn.functional as F
import torch, torch.distributed as dist import torch, torch.distributed as dist
@ -26,10 +25,9 @@ from utils import set_all_seed, print
from src.distributed.process_group_manager import setup_process_group_manager from src.distributed.process_group_manager import setup_process_group_manager
from src.parallel.pipeline_parallel import train_step_pipeline_1f1b, train_step_pipeline_afab, PipelineParallel from src.parallel.pipeline_parallel import train_step_pipeline_1f1b, train_step_pipeline_afab, PipelineParallel
from src.parallel.data_parallel.data_parallel_bucket import DataParallel from src.parallel.data_parallel.data_parallel_bucket import DataParallel
# from src.parallel.context_parallel import ContextParallel from src.parallel.context_parallel import ContextParallel
from model import LLaMA from model import Llama
import wandb import wandb
import multiprocessing
class MicroBatchDataLoader(DataLoader): class MicroBatchDataLoader(DataLoader):
def __init__(self, global_batch_size, micro_batch_size, seq_length, dataset_name, tokenizer_name, grad_acc = 1, split="train", num_samples=None, num_workers=0): def __init__(self, global_batch_size, micro_batch_size, seq_length, dataset_name, tokenizer_name, grad_acc = 1, split="train", num_samples=None, num_workers=0):
@ -205,9 +203,7 @@ if __name__ == "__main__":
config.num_attention_heads = 16 config.num_attention_heads = 16
config.num_key_value_heads = 4 config.num_key_value_heads = 4
model = LLaMA( model = Llama(config=config)
config=config
)
if pgm.process_group_manager.global_rank == 0 and args.use_wandb: if pgm.process_group_manager.global_rank == 0 and args.use_wandb:
wandb.init( wandb.init(
@ -231,13 +227,14 @@ if __name__ == "__main__":
TensorParallel(model) TensorParallel(model)
# if pgm.process_group_manager.cp_size > 1: # if pgm.process_group_manager.cp_size > 1:
# model = ContextParallel(model, config) #TODO: do at the very end when we have fix convergence issue
# model = ContextParallel(model, config)
if pgm.process_group_manager.pp_world_size > 1: if pgm.process_group_manager.pp_world_size > 1:
model = PipelineParallel(model, config) model = PipelineParallel(model, config)
if pgm.process_group_manager.dp_world_size > 1: if pgm.process_group_manager.dp_world_size > 1:
model = DataParallel(model, pgm.process_group_manager.dp_group) model = DataParallel(model)
model.to(device) model.to(device)
model.train() model.train()
@ -267,6 +264,7 @@ if __name__ == "__main__":
# average the loss across all DP/CP ranks # average the loss across all DP/CP ranks
if pgm.process_group_manager.dp_world_size > 1 or pgm.process_group_manager.cp_world_size > 1: if pgm.process_group_manager.dp_world_size > 1 or pgm.process_group_manager.cp_world_size > 1:
#TODO: use all_reduce function from distributed_primitives.py
loss_tensor = torch.tensor([loss], dtype=torch.float32, device=device) loss_tensor = torch.tensor([loss], dtype=torch.float32, device=device)
handle = dist.all_reduce(loss_tensor, group=pgm.process_group_manager.cp_dp_group, async_op=True, op=dist.ReduceOp.AVG) handle = dist.all_reduce(loss_tensor, group=pgm.process_group_manager.cp_dp_group, async_op=True, op=dist.ReduceOp.AVG)