Repository navigation
[TorchAO] MX training native PyTorch on XPU #2326
Description
🚀 The feature, motivation and pitch
The scaled_mm(pytorch/pytorch#165978) and inference(pytorch/ao#3248) is WIP, You can enable the realted training UTs with emulate path and comments your PR in pytorch/ao#2917 to help us to track the total status.
Reference: https://github.com/pytorch/ao/tree/main/torchao/prototype/mx_formats
Alternatives
No response
Additional context
No response
Activity
[-][TorchAO] MX training native PyTorch[/-][+][TorchAO] MX training native PyTorch on XPU[/+]karol-brejna-i commentedon Apr 14, 2026
Progress tracked in: https://jira.devtools.intel.com/browse/PYTORCHDGQ-7740
xiaowangintel commentedon May 28, 2026
There are three draft PRs(pytorch/pytorch#181728, pytorch/pytorch#181727 and pytorch/pytorch#181726) of pytorch related to MXTensor here. You can use them to verify the functionality of mxtensor.
ugolowic commentedon Jun 12, 2026
@xiaowangintel , @karol-brejna-i
I applied the PRs, built pytorch and ran ao UTs. All expected tests pass, including the ones testing matrix multiplication.
daisyden commentedon Nov 11, 2026
label-issue: #2326
Root cause: not a defect. Feature request to enable TorchAO MX-format (microscaling)
training on native PyTorch XPU via the emulate path. No failure or test case reported.
typeFeaturetesttest: utmodulemodule: aopriorityLowApplied to this issue. Generated by a local
label-issuerun.