Skip to content

[apex FusedAdam] crash workaround #249

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

Merged
stas00 merged 5 commits into from
Feb 18, 2022
Merged

[apex FusedAdam] crash workaround #249

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
57398b8
[apex FusedAdam] crash workaround
stas00 Feb 18, 2022
8eedd3c
Trigger CI
stas00 Feb 18, 2022
6609ce6
Trigger CI
stas00 Feb 18, 2022
498d868
fix
stas00 Feb 18, 2022
224443b
fix
stas00 Feb 18, 2022
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
15 changes: 14 additions & 1 deletion megatron/optimizer/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,7 +44,20 @@ def _get_params_for_weight_decay_optimization(modules):
[p for n, p in list(module_._parameters.items())
if p is not None and n == 'bias'])

return weight_decay_params, no_weight_decay_params
# XXX: temp hack to workaround the crash in apex FusedAdam's multi_tensor_applier
#
# it crashes when the param count is larger than a certain size which we hit at 200B over 80
# A100 gpus - I think around 2.7B per gpu, so halving it works around the issue
param_count = len(weight_decay_params['params'])
first_half = weight_decay_params['params'][:param_count // 2]
second_half = weight_decay_params['params'][param_count // 2:]

first_half = { 'params': first_half }
second_half = { 'params': second_half }

return first_half, second_half, no_weight_decay_params

#return weight_decay_params, no_weight_decay_params


def get_megatron_optimizer(model):
Expand Down