Skip to content

v0.111.0

Choose a tag to compare

@ooples ooples released this 07 Jul 15:35
230fbe9

0.111.0 (2026-07-07)

Features

  • simd: dual-path Adam kernel — bit-exact default + AIDOTNET_FAST_MATH FMA fast path (#738) (d670ce3)

Bug Fixes

  • autodiff: make 2D AffineGrid differentiable w.r.t. theta (#747) (72e4116)
  • bench: gate gemm-bench on the median per-round paired ON/OFF ratio (#741) (39194bb)
  • gpu: fused-optimizer weight-coherence — refresh resident device buffer after in-place update (#739) (8e61dc4)
  • gpu: hardware-independent determinism for scatter-add grad + cuBLAS under SetDeterministicMode (Closes #742) (#744) (b815872)

Performance

  • gemm: beta=0 write-first path — skip redundant C-zeroing on pure C:=A·B (#743) (2aac704)
  • kill training spin-storm — lower resident-pool MRES spin (2047 -> 32) (#736) (8ede01f)
  • zero-alloc training via TensorArena reuse + shape-safe ring (DRAFT) (#734) (3f7178a)