osmarks/nanogpt-experiments
1
0
Fork 0
mirror of https://github.com/osmarks/nanogpt-experiments.git synced 2024-12-18 06:00:29 +00:00
Code Issues Releases Wiki Activity

Merge pull request #225 from otaviogood/grad_accum

Browse Source 
Fix for gradient_accumulation_steps training slow
This commit is contained in:
Andrej 2023-04-17 20:11:25 -07:00 committed by GitHub
commit 21f9bff7e4
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 9 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
config/train_gpt2.py
View File

@ -10,7 +10,7 @@ wandb_run_name='gpt2-124M'
# 12 batch size * 1024 block size * 5 gradaccum * 8 GPUs = 491,520 # 12 batch size * 1024 block size * 5 gradaccum * 8 GPUs = 491,520
batch_size = 12 batch_size = 12
block_size = 1024 block_size = 1024
gradient_accumulation_steps = 5 gradient_accumulation_steps = 5 * 8
# this makes total number of tokens be 300B # this makes total number of tokens be 300B
max_iters = 600000 max_iters = 600000

1
config/train_shakespeare_char.py
View File

@ -14,6 +14,7 @@ wandb_project = 'shakespeare-char'
wandb_run_name = 'mini-gpt' wandb_run_name = 'mini-gpt'
dataset = 'shakespeare_char' dataset = 'shakespeare_char'
gradient_accumulation_steps = 1
batch_size = 64 batch_size = 64
block_size = 256 # context of up to 256 previous characters block_size = 256 # context of up to 256 previous characters

10
train.py
View File

@ -45,7 +45,7 @@ wandb_project = 'owt'
wandb_run_name = 'gpt2' # 'run' + str(time.time()) wandb_run_name = 'gpt2' # 'run' + str(time.time())
# data # data
dataset = 'openwebtext' dataset = 'openwebtext'
gradient_accumulation_steps = 5 # used to simulate larger batch sizes gradient_accumulation_steps = 5 * 8 # used to simulate larger batch sizes
batch_size = 12 # if gradient_accumulation_steps > 1, this is the micro-batch size batch_size = 12 # if gradient_accumulation_steps > 1, this is the micro-batch size
block_size = 1024 block_size = 1024
# model # model
@ -84,16 +84,20 @@ if ddp:
init_process_group(backend=backend) init_process_group(backend=backend)
ddp_rank = int(os.environ['RANK']) ddp_rank = int(os.environ['RANK'])
ddp_local_rank = int(os.environ['LOCAL_RANK']) ddp_local_rank = int(os.environ['LOCAL_RANK'])
ddp_world_size = int(os.environ['WORLD_SIZE'])
device = f'cuda:{ddp_local_rank}' device = f'cuda:{ddp_local_rank}'
torch.cuda.set_device(device) torch.cuda.set_device(device)
master_process = ddp_rank == 0 # this process will do logging, checkpointing etc. master_process = ddp_rank == 0 # this process will do logging, checkpointing etc.
seed_offset = ddp_rank # each process gets a different seed seed_offset = ddp_rank # each process gets a different seed
assert gradient_accumulation_steps % torch.cuda.device_count() == 0
gradient_accumulation_steps //= torch.cuda.device_count()
else: else:
# if not ddp, we are running on a single gpu, and one process # if not ddp, we are running on a single gpu, and one process
master_process = True master_process = True
seed_offset = 0 seed_offset = 0
gradient_accumulation_steps *= 8 # simulate 8 gpus ddp_world_size = 1
print("total number of tokens per iteration:", batch_size * block_size * gradient_accumulation_steps) tokens_per_iter = gradient_accumulation_steps * ddp_world_size * batch_size * block_size
print(f"tokens per iteration will be: {tokens_per_iter:,}")
if master_process: if master_process:
os.makedirs(out_dir, exist_ok=True) os.makedirs(out_dir, exist_ok=True)