Skip to content

Remove abs from save list for per op AC to fix bug causing unexpected increase in peak memory usage for float8 training with row-wise scaling #820

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 4 commits into from
Feb 6, 2025
Merged

Remove abs from save list for per op AC to fix bug causing unexpected increase in peak memory usage for float8 training with row-wise scaling #820

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
c6e9d64
remove abs from save list for per op ac
danielvegamyhre Feb 4, 2025
d607298
update comment
danielvegamyhre Feb 5, 2025
730592a
trigger build
danielvegamyhre Feb 5, 2025
838ebab
trigger ci
danielvegamyhre Feb 6, 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/parallelisms/parallelize_llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -217,8 +217,8 @@ def apply_tp(
torch.ops.aten._scaled_dot_product_flash_attention.default,
torch.ops._c10d_functional.reduce_scatter_tensor.default,
# for low precision training, it's useful to always save
# the result of max(abs(tensor))
torch.ops.aten.abs.default,
# the result of max, since the absolute maximum is
# used to compute the scaling factor for quantization.
torch.ops.aten.max.default,
Copy link
Contributor

@tianyu-l tianyu-l Feb 5, 2025

Choose a reason for hiding this comment

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

since abs needs to be recomputed, do we still want to keep max? Asking in another way, why do we need to keep abs in the first place, since we already keep the result of max? @vkuzo

Copy link
Contributor

@vkuzo vkuzo Feb 5, 2025

Choose a reason for hiding this comment

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

we don't need to keep abs, I made a mistake adding it in my original PR. We just didn't see the cost of the mistake until the rowwise scaled float8 recipe.

}

Expand Down
Loading