cortiq 0.8.44

[0.8.44] — 2026-10-29

Fixed

• wgpu on the AMD / Intel GPUs of Intel Macs (Metal): naga's workgroup-memory
  zeroing, injected at the top of every kernel, cost 0.5–2 ms a dispatch there
  (an empty kernel takes 7 µs). Every pipeline now takes its compile options
  from one rule (gpu_wgpu::pco): Metal on a non-Apple GPU skips the zeroing,
  every other device keeps it; CMF_WGPU_NOZERO=1|0 forces either way. The
  kernels write their workgroup arrays before reading them: the zeroed and the
  plain run write the same greedy text, the wgpu inference tests pass without
  it, and ppl matches the CPU. Spark-X2.6-4B q4tp on a Radeon Pro 560X
  (MacBook Pro 2019): decode 1.1 → 13.6 tok/s, prompt 10.9 → 19.2 tok/s.

Changed

• AMD GPUs under Metal decode q4tp through wave64 matvecs (gpu_wgpu::amd_mv):
  masked-nibble dequant, shuffle reductions, no ladder table or barrier —
  sixteen rows a wave from 8192 rows, four below; the pair and fused gate+up
  kernels stand down there. Radeon Pro 560X: GPU time 64 → 55 ms a token,
  decode 13.9 → 15.9 tok/s (the i9 CPU of the same Mac: ~10). Not the CPU's
  add order (|Δ|/max ≈ 2e-6). CMF_AMD_MV=0 keeps the previous kernels.

Added

• wgpu_q4tp_mv_variants (ignored test): q4tp decode matvec candidates from
  a WGSL file against q4tp_matvec16w on device-local weights, per shape.