v0.4.1
·
34 commits
to main
since this release
Immutable
release. Only release title and notes can be modified.
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