[XPU][Tracker] Enable `torch._scaled_matmul`
Open
module: xpu
triaged
- Dominant language
- Python
- Stars
- 103k
- Forks
- 29.5k
- PR merge metrics
- PR metrics pending
Description
This is a task tracker for enabling `torch._scaled_matmul()` and related ops on XPU devices.
# Tasks in pytorch repo
- [x] Basic Implementation: Merge in the basic scaling. (Tensorwise / Rowwise) https://github.com/pytorch/pytorch/pull/165978
- [x] Align API with `torch._scaled_mm_v2`. https://github.com/pytorch/pytorch/pull/167518
- [x] Op Registration: https://github.com/pytorch/pytorch/pull/166056
- [ ] Support more scalings and datatypes (NVFP4 / MXFP8 /block-wise scaling).
# Tasks in torch libraries
- [x] Integrating the path in torchao
cc @gujinghui @EikanWang @fengyuan14 @guangyey
Contributor guide
Assessment
This issue has not been assessed yet.