Skip to content

Fixes accuracy error for mxfp8 linear #3357

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
vkuzo merged 2 commits into from
Nov 24, 2025
Merged

Fixes accuracy error for mxfp8 linear #3357

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4848521
Fixes accuracy error for mxfp8 linear
syed-ahmed Nov 20, 2025
3c7a8e8
Adds todo
syed-ahmed Nov 20, 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
5 changes: 4 additions & 1 deletion torchao/prototype/mx_formats/kernels.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -881,7 +881,10 @@ def _get_mxfp8_quant_autotune_configs():
# can be improved in the future.
results = []
for ROW_TILE_SIZE in (128, 256, 512):
for COL_TILE_SIZE in (128, 256, 512):
# TODO: we can't use 512 for COL_TILE_SIZE.
# This is likely a triton bug, tracked in
# https://github.com/pytorch/ao/issues/3362
for COL_TILE_SIZE in (128, 256):
for num_warps in (4, 8):
for num_stages in (2, 3):
config = triton.Config(
Expand Down
Loading