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

Do not aggregate the losses since last log step #779

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
+9 −11
Open

Do not aggregate the losses since last log step #779

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
726efc4
Do not aggregate the losses since last log step
carmocca Jan 7, 2025
7a3e1fb
Dtensor
carmocca Jan 9, 2025
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
4 changes: 2 additions & 2 deletions torchtitan/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,12 +34,12 @@ def get_device_info():
device_type, device_module = get_device_info()


def dist_max(x: Union[int, float], mesh: DeviceMesh) -> float:
def dist_max(x: Union[int, float, torch.Tensor], mesh: DeviceMesh) -> float:
tensor = torch.tensor(x).to(device_type)
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

AFAIK these will result in no-ops if a tensor of with the same device type is passed.

return funcol.all_reduce(tensor, reduceOp=c10d.ReduceOp.MAX.name, group=mesh).item()


def dist_mean(x: Union[int, float], mesh: DeviceMesh) -> float:
def dist_mean(x: Union[int, float, torch.Tensor], mesh: DeviceMesh) -> float:
tensor = torch.tensor(x).to(device_type)
return funcol.all_reduce(tensor, reduceOp=c10d.ReduceOp.AVG.name, group=mesh).item()

Expand Down
16 changes: 7 additions & 9 deletions train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,7 @@
import torch

from torch.distributed.elastic.multiprocessing.errors import record
from torch.distributed.tensor import DTensor

from torchtitan import utils
from torchtitan.checkpoint import CheckpointManager, TrainState
Expand Down Expand Up @@ -228,7 +229,6 @@ def loss_fn(pred, labels):
)

# variables used to keep info for metrics logging
losses_since_last_log = []
ntokens_since_last_log = 0
data_loading_times = []
time_last_log = time.perf_counter()
Expand Down Expand Up @@ -327,26 +327,25 @@ def loss_fn(pred, labels):
# it issues a single all-reduce for all parameters at once for better performance
float8_handler.precompute_float8_dynamic_scale_for_fsdp(model_parts)

losses_since_last_log.append(loss)

# log metrics
if (
train_state.step == 1
or train_state.step % job_config.metrics.log_freq == 0
):
losses = [loss.item() for loss in losses_since_last_log]
avg_loss, max_loss = sum(losses) / len(losses), max(losses)
if (
parallel_dims.dp_replicate_enabled
or parallel_dims.dp_shard_enabled
or parallel_dims.cp_enabled
):
loss = loss.detach()
if isinstance(loss, DTensor):
loss = loss.full_tensor()
global_avg_loss, global_max_loss = (
utils.dist_mean(avg_loss, world_mesh["dp_cp"]),
utils.dist_max(max_loss, world_mesh["dp_cp"]),
utils.dist_mean(loss, world_mesh["dp_cp"]),
utils.dist_max(loss, world_mesh["dp_cp"]),
Comment on lines +344 to +345
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

These use functional collectives under the hood so there shouldn't be issues with passing in the same tensor reference

Comment on lines +344 to +345
Copy link
Contributor

@tianyu-l tianyu-l Jan 7, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It seems with Tensor Parallel, the loss is a DTensor, which doesn't support functional collectives. Also we should not require gradients on this all-reduce.
Maybe it's still fine to do .item() outside as before? or use detach and full_tensor() ?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I added the detach and full_tensor calls

)
else:
global_avg_loss, global_max_loss = avg_loss, max_loss
global_avg_loss = global_max_loss = loss.item()

# update train state
train_state.log_steps.append(train_state.step)
Expand Down Expand Up @@ -396,7 +395,6 @@ def loss_fn(pred, labels):
f"{color.magenta}mfu: {mfu:.2f}%{color.reset}"
)

losses_since_last_log.clear()
ntokens_since_last_log = 0
data_loading_times.clear()
time_last_log = time.perf_counter()
Expand Down