release: 0.8.8 — glq-quantize defaults to the trellis codebook

[cnygaard]

The headline is a behaviour change: --codebook now defaults to trellis (3INST variant) instead of e8_shell. Anyone who quantizes without naming a codebook gets a different format from this release on.

That was already the recommended path — the README said so, and the checkpoint table marked a trellis build as "fastest GLQ decode" — while the CLI still handed out lattice checkpoints. Trellis wins where bits are scarcest: SmolLM3-3B at 2 bpw is PPL 11.94 vs 13.79.

What a user notices

• glq-quantize produces trellis-3INST by default. Uniform integer bit-rates only (2–8); fractional and per-layer mixed precision now need --codebook e8_shell, and the refusal says so.
• The variant moved too — hyb → 3inst, because the fused CUDA kernels consume a 3INST checkpoint. Defaulting to trellis alone would have shipped checkpoints with no fast path.
• Pre-flight names the gcc/nvcc fix instead of letting the build die in ninja. fedora:44 ships gcc 16 against CUDA's cap of 15; a compat compiler plus NVCC_CCBIN makes a source install work there for the first time. It is a note, not a blocker — prebuilt wheels compile nothing and are unaffected.
• The README documents the Blackwell/FlashInfer trap: on sm_120 without a CUDA Toolkit, vLLM's sampler compiles at startup and takes the engine down. glq-chat falls back on its own; vllm serve and LLM(...) do not.

Version pins

Five places: pyproject.toml, glq/__init__.py, the Dockerfile's GLQ_VERSION default plus its three build comments, and the glq[quantize] pin in infra/setup.sh.tftpl. install.sh needs no change — it defaults to latest on PyPI.

The remaining 0.8.7 strings in .github/workflows/docker.yml and tests/test_installer_distros.py are prose about past releases, not pins, and are deliberately left.

Validation

• Full local suite on the trellis-default change: 1085 passed, 882 skipped, 0 failed.
• The pre-flight advice was validated end to end on real fedora:44: follow the package advice, then the compat-compiler advice, then install.sh --glq-source with NVCC_CCBIN exported → install exit 0, EXT_OK:True, 0 gcc errors. The extension builds where it previously could not.
• Distro matrix (11 distros, RTX PRO 6000): pre-flight advice correct and sufficient and GPU visible on all 11, including arch, azurelinux, opensuse and amazon2023 which had never been executed. Ubuntu 26.04 passes both install modes.
• A matrix run against this release commit is in flight; results will be added before merge.

Existing checkpoints

Unaffected. Nothing here changes the encode or storage format of an already-quantized model — only which codebook a new quantization picks when the user does not choose.