Skip to content

Fix gradient explosion in bfloat16+float8 rowwise quantization #3171

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

Closed
wants to merge 1 commit into from
+3 −0
Closed

Fix gradient explosion in bfloat16+float8 rowwise quantization #3171

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
3 changes: 3 additions & 0 deletions torchao/float8/float8_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,6 +66,9 @@ def tensor_to_amax(
else:
assert scaling_granularity is ScalingGranularity.AXISWISE, "unsupported"
assert axiswise_dim is not None, "unsupported"
# Fix for bfloat16 precision amplification in amax_to_scale()
if x.dtype == torch.bfloat16:
x = torch.abs(x).to(torch.float32)
amax = torch.amax(torch.abs(x), dim=axiswise_dim, keepdim=True)

# If the user asked for distributed reduction, do it.
Expand Down
Loading