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

Bugfixes: Grad norm scaling in TP #2436

Draft
wants to merge 6 commits into
base: main
Choose a base branch
from
Draft

Bugfixes: Grad norm scaling in TP #2436

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
0688528
Use implicit replication in distributed train step to fix grad norm c…
nathan-az Feb 25, 2025
8f4c734
calculate and report grad norm even if not clipping, fix tok/sec repo…
nathan-az Feb 25, 2025
136d1ce
fix foreach typo
nathan-az Feb 25, 2025
8d0c044
Adjust grad norm scaling and tps counter
nathan-az Feb 25, 2025
d498a58
Inline passing params to grad norm calc
nathan-az Feb 25, 2025
71daab1
Add None filter
nathan-az Feb 25, 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
45 changes: 32 additions & 13 deletions recipes/full_finetune_distributed.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,7 @@
init_process_group,
)
from torch.distributed._tensor import DTensor
from torch.distributed.tensor.experimental import implicit_replication
from torch.distributed.tensor.parallel import parallelize_module
from torch.optim import Optimizer
from torch.utils.data import DataLoader, DistributedSampler
Expand Down Expand Up @@ -818,6 +819,8 @@ def train(self) -> None:
torch.distributed.all_reduce(running_loss)

# We multiply by world_size to undo FSDP2 gradient normalization.
# TODO: confirm whether below should divide by `tensor_parallel_dim` to
# adjust grad norm for overcounting tokens from the all-reduce
current_loss = current_loss * (self.world_size / num_tokens)

current_loss.backward()
Expand All @@ -830,16 +833,27 @@ def train(self) -> None:
# This will ensure that the logged loss matches what we're optimizing
torch.distributed.all_reduce(running_loss)
# Manually scale the gradients from unnormalized loss by total # of tokens
# We multiply by world_size to undo FSDP2 gradient normalization.
training.scale_grads(self._model, self.world_size / num_tokens)
if self._clip_grad_norm is not None:
grad_norm = torch.nn.utils.clip_grad_norm_(
self._model.parameters(),
max_norm=float(self._clip_grad_norm),
)
# If sharded, collect the DTensor here
if isinstance(grad_norm, DTensor):
grad_norm = grad_norm.full_tensor()
# We multiply by dp_size to undo FSDP2 gradient normalization.
training.scale_grads(self._model, self.dp_size / num_tokens)
with implicit_replication():
if self._clip_grad_norm is not None:
grad_norm = torch.nn.utils.clip_grad_norm_(
self._model.parameters(),
max_norm=float(self._clip_grad_norm),
)
else:
grad_norm = torch.nn.utils.get_total_norm(
[
p.grad
for p in self._model.parameters()
if p.grad is not None
],
norm_type=2,
error_if_nonfinite=False,
foreach=True,
)
if isinstance(grad_norm, DTensor):
grad_norm = grad_norm.full_tensor()
self._optimizer.step()
self._optimizer.zero_grad(set_to_none=True)

Expand Down Expand Up @@ -871,15 +885,20 @@ def train(self) -> None:
else self._optim_ckpt_wrapper
),
),
# secondary tp dim division adjusts for tp ranks share the tokens during training
# so the above all-reduce over-counts
"tokens_per_second_per_gpu": num_tokens
/ (time_per_step * self.world_size),
/ (
time_per_step
* self.world_size
* self.tensor_parallel_dim
),
"grad_norm": grad_norm,
}
if self._log_peak_memory_stats:
log_dict.update(
training.get_memory_stats(device=self._device)
)
if self._clip_grad_norm is not None:
log_dict.update({"grad_norm": grad_norm})
self._metric_logger.log_dict(
log_dict,
step=self.global_step,
Expand Down