Skip to content

v0.8.0 — parametric substrate

Choose a tag to compare

@afflom afflom released this 10 Jul 01:37
· 270 commits to main since this release

Hologram v0.8.0

Installation

Via Git Dependency

Add to your Cargo.toml:

[dependencies]
# Depend on the specific crates you need (all share one version):
hologram-compiler = { git = "https://github.com/Hologram-Technologies/hologram", tag = "v0.8.0" }
hologram-exec     = { git = "https://github.com/Hologram-Technologies/hologram", tag = "v0.8.0" }
hologram-backend  = { git = "https://github.com/Hologram-Technologies/hologram", tag = "v0.8.0" }

CLI Binary

Download the appropriate binary for your platform from the assets below and add it to your PATH.

Changes

Commits since v0.7.2

  • fix(backend,bench): dispatch the fused W8A32 decode kernel on x86; bench opts into W8A8 (50b1f41)
  • fix(compiler): reject a non-zero zero-point on the output-major path at compile time (e704f4a)
  • perf(backend): block the output-major integer GEMVs at m > 1; correct the claim (5f9fc16)
  • release: follow the deprecate-then-version-out lifecycle; bump to v0.8.0 (8ca84b6)
  • fix(compiler): validate the OUTPUT_MAJOR declaration at compile time, not load (149a39a)
  • perf(backend): pin + document the integer GEMV's m-scaling defect (5eccb0b)
  • refactor(backend)+docs: typed wasm GEMV job; re-capture small-m benchmarks (19b163a)
  • fix(compiler,exec): separate weight layout from W8A8 numerics; no silent-wrong (1588748)
  • perf(backend): fix row-index-dependent bytes; reach wasm/NEON small-m; widen low-R GEMV (175bb15)
  • perf(backend): batch the small-m remainder rows into one pass over B (6e9acec)
  • refactor+docs: centralize decode gates, demote internals, correct stale claims (e907cc2)
  • feat: weight-slot declaration — a weightless weight can reach the fast decode GEMV (14f591d)
  • perf(backend): NEON e8cb kernel + wasm/NEON bulk widen; gate wasm dead code (1844e0d)
  • feat(backend,compiler,archive): the codebook is model data, not engine data (7eb9876)
  • refactor(backend,compiler): QuantTier registry — one row per weight encoding (0dd11fb)
  • refactor(types,backend,compiler,exec): one canonical DTypeId; total accessors (96bd421)
  • test(archive): freeze op_signature + wire bytes for legacy calls (κ guard) (8fe87d0)

What's Changed

  • cleanup: parametric substrate — one dtype vocabulary, per-model codebooks, no silent-wrong (v0.8.0) by @afflom in #37

Full Changelog: v0.7.2...v0.8.0