Repository navigation

perf(nvfp4): faster mat-vec, split-k tiles and nax kernels - #95

Merged

asher merged 7 commits into

mainfrom

perf/nvfp4-matvec

Oct 18, 2026

Merged

perf(nvfp4): faster mat-vec, split-k tiles and nax kernels#95
asher merged 7 commits into
mainfrom
perf/nvfp4-matvec

Commits on Oct 18, 2026
- perf(nvfp4): mat-vec on the mxfp4 lane layout with word loads
  asher
  committed
  c69a986
- perf(nvfp4): split-k tiles and nax kernels
  asher
  committed
  c3187d0
- test(nax): nvfp4 outlier test uses weights in the other codecs' range
  asher
  committed
  501b7cc
- test(moe): nvfp4 in the nax gather tests
  asher
  committed
  cd5b723
- feat(bench): nvfp4 weights and a k-ceiling tile a/b in the nax routing benches
  asher
  committed
  fee91c4
- perf(nvfp4): nax routing rows and verify_nax split count measured on m5
  asher
  committed
  734ed5e
- chore(release): 0.5.2
  asher
  committed
  f275fa2