Skip to content

Tags: bledden/tridec

Tags

v0.2.2

Toggle v0.2.2's commit message
release: v0.2.2 — CUDA-graph fast path (default-on) + d>=15 BP ceilin…

…g lift + triton_msl rename

Bumps 0.2.1 -> 0.2.2 and finalizes the CHANGELOG "Unreleased" section. Three
GPU-path changes, all validated cross-vendor (H200 CUDA + MI300X ROCm7):
- #4 CUDA-graph fast path, default-on for small-batch min-sum BP
  (cuda_graph="auto", S<=256; ~1.9x batch-1 on H200)
- #6 surface d>=15 BP-kernel ceiling lifted (1-D grid flatten; d=15/d=17 now
  decode on CUDA+ROCm, was capped at d=14)
- triton_metal -> triton_msl rename (experimental Metal backend; upstream rebrand)
Relay-BP / accuracy paths byte-unchanged from 0.2.1.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

v0.2.1

Toggle v0.2.1's commit message
release: v0.2.1 — bump version + docs

- pyproject version 0.2.0 -> 0.2.1; description now names Metal alongside
  CUDA/ROCm (tri-platform).
- CHANGELOG: Unreleased -> 0.2.1 (2026-06-14); note CUDA/ROCm kernels are
  byte-identical to 0.2.0 (lift is Metal-only, auto-dispatch is routing) so
  the 0.2.0 receipts stand; re-confirmed no regression on H200/MI300X.
- CITATION 0.2.1 / 2026-06-14.
- README Status: 0.2.1 released; relay auto-dispatch validated on Metal +
  H200 (CUDA) + MI300X (ROCm); Wilson-CI statistical tier noted.

v0.2.1 = relay auto-dispatch (#5, megakernel default on GPU) + Wilson-CI
statistical gate (#1) + the Metal megakernel BLOCK lift. Validated on all
three platforms this cycle (tests + public-API dispatch + latency).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

v0.2.0

Toggle v0.2.0's commit message
release prep: v0.2.0 — megakernel docs (opt-in framing, #5), CHANGELO…

…G, version bump

Coda v0.2 review folded in: README v0.2 block (relay megakernel headline +
both honest negatives + per-arch autotune + Metal BLOCK=32 transient note),
batch-1 cross-vendor limit stated plainly (H200 2.47x ahead of MI300X),
issue #5 prominence (README/CHANGELOG/module docstring), Status -> 0.2.0.
BP megakernel KEPT (Coda keep-test: named use case = single-shot latency,
batch-1 1.7x over two-kernel BP). pyproject 0.1.0 -> 0.2.0.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

v0.1.0

Toggle v0.1.0's commit message
release prep: CHANGELOG + CITATION.cff + pre-tag review fixes

Coda-review items: install sharp-edge note (bare install is numpy-only),
SemVer-zero API-stability line, §3.1 carried-receipt vs packaged-revalidation
reading note, CHANGELOG leading with the 31/32-exact no-regression result and
the three reproducibility findings.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>