Skip to content

v0.4.1

Choose a tag to compare

@asher asher released this 27 Aug 03:45
· 34 commits to main since this release
Immutable release. Only release title and notes can be modified.
7480b39

What's Changed

  • feat(qwen4exp,sdpa): block-sparse qsa prefill and paged decode, hc_lowrank fusions, dsa 4-head band by @asher in #44
  • perf(matmul): vector grid loaders (iq1-3) and iq4 pair LUT for pre-M5 GPUs by @asher in #45
  • perf(moe): widen glu-gather nx16 threadgroup cap to cover MTP-verify widths by @asher in #47
  • perf(hc): lane-serial epilogue kernel, flat in verify width by @asher in #48

Full Changelog: v0.4.0...v0.4.1