v0.4.12

What's Changed

• feat(kernels): prismML's PQ2_0 and PTQ1_0 codecs with hadamard_rotate and fused rotation kernels by @asher in #72
• perf(qmv): static row loop, mxfp4 mat-vec lane layout, opt-in core-count-scaled fine tiling by @asher in #73

Full Changelog: v0.4.11...v0.4.12

v0.4.11

What's Changed

• fix(metal): router top-k keeps NaN rows in range by @asher in #70

Full Changelog: v0.4.10...v0.4.11

v0.4.10

What's Changed

• fix(gguf): zero-copy read support for >16GiB tensors by @asher in #67
• feat(dsa): DeepSeek-V4 sparse attention kernels, FP4 latent pool and V4.1 QAT forms by @asher in #68

Full Changelog: v0.4.9...v0.4.10

v0.4.9

What's Changed

• feat(sdpa): per-row ends on the decode kernels for ragged batches by @asher in #65

Full Changelog: v0.4.8...v0.4.9

v0.4.8

What's Changed

• feat(kernel): KDA chunked recurrence and prefill glue, sorted-MoE mix, fused hc cycle, indexed SDPA by @asher in #63

Full Changelog: v0.4.7...v0.4.8

v0.4.7

What's Changed

• fix(arena): mmap the arena buffers so a release returns pages to the kernel by @asher in #61
• fix(build): cleanup warnings by @asher in #60

Full Changelog: v0.4.6...v0.4.7

v0.4.6

What's Changed

• feat(kvarn): KVarN variance-normalized KV cache ops and SDPA kernels by @asher in #58

Full Changelog: v0.4.5...v0.4.6

v0.4.5

What's Changed

• perf(kernel): faster IQ gather decode, sigmoid router arm, q5_k shared-expert codec by @asher in #56

Full Changelog: v0.4.4...v0.4.5

v0.4.4

What's Changed

• feat(kernel): stq1_0 support, a 1.3125 bpw structured-sparse ternary codec by @asher in #54
• fix(gguf): release the GIL during load_gguf parse by @asher in #53

Full Changelog: v0.4.3...v0.4.4

v0.4.3

What's Changed

• fix(dsa): deterministic top-k tie admission (stable compaction, was racing atomics) by @asher in #52

Full Changelog: v0.4.2...v0.4.3