v0.7.2 — M=1 decode latency elimination

Hologram v0.7.2

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 = "v0.7.2" }
hologram-exec     = { git = "https://github.com/Hologram-Technologies/hologram", tag = "v0.7.2" }
hologram-backend  = { git = "https://github.com/Hologram-Technologies/hologram", tag = "v0.7.2" }

CLI Binary

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

Changes

Commits since v0.7.1

• chore: bump version to v0.7.2 (25e6721)
• fix(backend): debug-only i8 overflow in e8cb test + refresh API snapshot (49fd47b)
• test(backend): harden the pool barrier + E8-codebook coverage (943e387)
• perf(backend): wasm-threads fork-join for the E8-codebook decode GEMV (9bfe613)
• perf(backend): spin-then-park pool + atomic barrier — kill fork/join latency (9c9e2e2)
• perf(backend,compiler): E8 lattice-codebook (VQ) decode tier — 1 bit/weight (8514632)
• perf(backend): multi-thread the omajor W8A8 int8 decode GEMV (native) (30c4bcd)
• perf(backend): streamed bf16/f16 GEMV — kill per-token weight re-materialization (07f61e7)

What's Changed

• perf: eliminate the M=1 decode substrate bottleneck (bf16 streamed GEMV + threaded int8) by @afflom in #36

Full Changelog: v0.7.1...v0.7.2