Skip to content

v0.7.1

v0.7.1: chore: Release mlx-native version 0.7.1
@robertelee78 robertelee78 tagged this 09 May 02:41 
Three ADR-020 ac#5 commits since v0.7.0. All additive; no API breaks.

- iter-A (d0de92a): port packed-U32 dense affine matmul kernel
  (qmm_affine_t_packed) for MLX-format weights with U32-packed
  payloads.
- iter-E (de0bcb9): relax dispatch buffer validation on the
  qmm_affine_t_packed path — accept the packed-U32 layout that
  iter-A introduced without false-positive shape rejections.
- iter-C2.3 (51cfca9): quantized_matmul_id_into variant — caller-
  provided output buffer signature for the MoE expert-routed
  matmul, enabling in-place output writes without the allocate-
  then-copy pattern.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Assets 2
Loading