This release includes 1 security fix for security teams reviewing exposed deployments.
Topics
Affected surfaces
Summary
AI summaryIntroduces the 3INST trellis decode variant with arithmetic in‑kernel decoding and fixes an upstream idle‑warp out‑of‑bounds read.
Full changelog
Trellis 3INST — lookup-free trellis decode
The trellis (TCQ) codebook gains the 3INST variant: each 16-bit trellis state is decoded arithmetically in-kernel (a hash + two fp16 halves) instead of through HYB's 512-entry shared-memory lookup table. Quality measured equal-or-slightly-better than HYB on SmolLM3-3B (SQNR 22.06 vs 21.1, wikitext-2 PPL 9.2425 vs 9.2454; AIME-2026 thinking 43.3% vs 40.4% avg@8, none significant), with a simpler and faster kernel path.
<R, IS_3INST>templated matvec/matmul kernels, zero dynamic shared memory- Select at quantize time with
GLQ_TRELLIS_VARIANT=3inst(--codebook trellis --bpw {2,3,4}) - Kernel storage layout: trellis checkpoints now store indices decoder-native (
trellis_layout: "kernel"); loading them requires glq ≥ 0.7.0 — older versions abort with a layout error instead of decoding garbage - Serves on vLLM 0.25 under full CUDA-graph capture; HF eager path included
- Fixes an upstream idle-warp out-of-bounds read in the trellis matmul preload
- Public checkpoint: xv0y5ncu/SmolLM3-3B-trellis-3inst-4bpw-kernel
Breaking Changes
- Treillis checkpoints now use decoder‑native index layout (`trellis_layout: "kernel"`); requires glq ≥ 0.7.0, older versions abort with a layout error
Security Fixes
- Fixes an upstream idle‑warp out‑of‑bounds read in the trellis matmul preload
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About Glq
All releases →Beta — feedback welcome: [email protected]