Repository navigation

Packed-panel CQ4 weight format: NEON + SME2 kernels - #709

Open

ncylich wants to merge 5 commits into

gemv-dispatchfrom

panel-format-neon

Open

Packed-panel CQ4 weight format: NEON + SME2 kernels#709
ncylich wants to merge 5 commits into
gemv-dispatchfrom
panel-format-neon

Commits on Jun 11, 2026
- Add --torch-dtype to convert for bf16-safe model loading
  ncylich
  committed
  115a257
- Disable the dense-MLP TQ fusion by default
  ncylich
  committed
  df4b2ac
- Add packed-panel CQ4 weight format and NEON kernels
  ncylich
  committed
  842e644
- Report converted tensor bytes from the written panel path
  ncylich
  committed
  15f9b98
- Add SME2 streaming kernels over the packed-panel weight format
  ncylich
  committed
  22df501