This release adds 2 notable features for engineering teams evaluating rollout.
✓ No known CVEs patched in this version
Topics
Summary
AI summaryFixed mixed‑precision profiling KeyError and added vLLM 0.25 support.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Performance | Medium |
Improves odd‑bpw decode speed to near parity with even‑bpw. Improves odd‑bpw decode speed to near parity with even‑bpw. Source: llm_adapter@2026-07-13 Confidence: high |
— |
| Bugfix | Medium |
Fixes KeyError in mixed‑precision profiling for e8p codebook. Fixes KeyError in mixed‑precision profiling for e8p codebook. Source: llm_adapter@2026-07-13 Confidence: high |
— |
Full changelog
Fast odd-bpw (3/5/7) decode
The trailing E81B residual stage now uses SM-saturating split-K GEMM/GEMV kernels (+ M-parallel grouped-MoE) with a smem-resident codebook, replacing the 1-warp lookup-matmul — odd-bpw decode was ~7–8× slower than even-bpw, now near parity. Bit-exact + deterministic; existing odd-bpw checkpoints work unchanged.
vLLM 0.25 support
glq_vllm detects MoE layers via RoutedExperts (0.25 split FusedMoE into a factory function), version-guarded for ≤0.23. Serves coherent across 26B-A4B MoE, SmolLM2/3, and Gemma-4 E4B/12B/31B.
Fixes
Mixed-precision (--min-bpw/--max-bpw) profiling KeyError on the e8p codebook.
pip install glq
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]