Skip to content

float8 rowwise training: add FSDP workaround #1629

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
merged 1 commit into from
Jan 31, 2025
Merged

float8 rowwise training: add FSDP workaround #1629

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
9 changes: 9 additions & 0 deletions torchao/float8/float8_linear.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -159,6 +159,15 @@ def backward(ctx, grad_output):
elif c.cast_config_weight_for_grad_input.scaling_type is ScalingType.DISABLED:
weight_t_maybe_fp8_dim0 = weight_hp_t
else:
if (
c.cast_config_weight_for_grad_input.scaling_granularity
is ScalingGranularity.AXISWISE
):
# workaround from https://github.com/pytorch/pytorch/issues/141881
# to avoid saving float8 weight from forward to backward when
# FSDP is on
weight_hp_t = weight_hp_t + (grad_output_reshaped[0, 0] * 0)

# Note: we need https://github.com/pytorch/pytorch/issues/136267
# to be solved to have a chance to reuse max(abs(weight, dim=...))
# from the forward to get max(abs(weight)) here without reading
Expand Down
Loading