Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

get distributed backend name via accelerator and check loss_scale before writing to tb #374

Merged
merged 4 commits into from
May 13, 2024
Merged

get distributed backend name via accelerator and check loss_scale before writing to tb #374

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
0ff1235
check loss_scale before writing to tb
polisettyvarma Apr 3, 2024
c3f76ee
get distributed backend name via accelerator
polisettyvarma Apr 3, 2024
33ba922
Merge branch 'microsoft:main' into any_device
polisettyvarma May 11, 2024
5cf2b1c
add hccl distributed backend support
polisettyvarma May 11, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion megatron/arguments.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1138,7 +1138,7 @@ def _add_distributed_args(parser):
help='overlap pipeline parallel communication with forward and backward chunks',
dest='overlap_p2p_comm')
group.add_argument('--distributed-backend', default='nccl',
choices=['nccl', 'gloo', 'ccl'],
choices=['nccl', 'gloo', 'ccl', 'hccl'],
help='Which backend to use for distributed training.')
group.add_argument('--distributed-timeout-minutes', type=int, default=10,
help='Timeout minutes for torch.distributed.')
Expand Down
2 changes: 1 addition & 1 deletion megatron/initialize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -222,7 +222,7 @@ def _initialize_distributed():
else:
if not torch.distributed.is_initialized():
torch.distributed.init_process_group(
backend=args.distributed_backend,
backend=get_accelerator().communication_backend_name(),
world_size=args.world_size, rank=args.rank,
timeout=timedelta(minutes=args.distributed_timeout_minutes))

Expand Down
2 changes: 1 addition & 1 deletion megatron/training.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -887,7 +887,7 @@ def training_log(loss_dict, total_loss_dict, learning_rate, iteration,
args.consumed_train_samples)
writer.add_scalar(f"lm-loss-training/{key}" + ' vs tokens', loss_dict[key],
args.consumed_train_tokens)
if args.fp16 and args.log_loss_scale_to_tensorboard:
if args.fp16 and loss_scale and args.log_loss_scale_to_tensorboard:
writer.add_scalar('loss-scale/loss-scale', loss_scale, iteration)
writer.add_scalar('loss-scale/loss-scale vs samples', loss_scale,
args.consumed_train_samples)
Expand Down