gridbook 0.3.0

gridbook 0.3.0

This release incorporates reviewed contributions from Jason Wong (@jsconsultancy), originally supplied as the JW2026 patch set, together with integration fixes and expanded coverage.

• Aligns ignore matching with compressed-tensors and verifies external codebook provenance before loading.
• Rejects malformed codebooks and incompatible native extensions instead of silently degrading.
• Adds single-launch CUDA FP4 activation QDQ, measured at 15.7–26.3× operator-level speedups during contribution validation.
• Adds an opt-in shared-memory grouped-MoE decode kernel. Jason's Laguna release measured 24.993 versus 23.585 tok/s on GB10, a 5.97% end-to-end improvement.
• Adds explicit, byte-budgeted FP4 stock prefill and stronger extension-cache diagnostics.
• Documents audited GB10 delegated-NVFP4 backend behavior.

The v2 decode kernel remains an explicit opt-in; inherited production dispatch is unchanged. GB10 cc12.1 is validated; cc12.0 remains unexecuted.

Major contributions: Jason Wong (@jsconsultancy) — #3, #4, #6, #7, #8, #9, #10, and #11. Review, integration, test expansion, and release validation: Robert Tand.

Full Changelog: v0.2.0...v0.3.0

gridbook 0.2.0

Full Changelog: v0.1.1...v0.2.0

gridbook 0.1.1

Full Changelog: v0.1.0...v0.1.1

gridbook 0.1.0

First release in which pip install gridbook produces a working plugin.

pip install gridbook   # into an environment that already has vLLM + torch

Why this is the first usable release

Before this, the CUDA sources lived outside the Python package while cuda_ext resolved them repo-root-relative, so any non-editable install found no kernels and fell silently to the correctness-only Triton path — one stderr warning as the only signal. The sources now ship inside the package (gridbook/csrc) as package-data. Reported from the field in #1.

Serving fixes (#1)

• Dense CB weights are resident once, not twice. process_weights_after_loading built a padded copy and left the original parameter live beside it. On the published Qwen3.6-27B artifact (21.38 GiB on disk): model weights 35.86 → 20.82 GiB, KV cache 57.09 → 72.04 GiB, KV capacity 1.40M → 1.76M tokens. Invisible on a 128 GB box, fatal on a 32 GB card.
• Fused-module scheme lookup is namespace-robust. Probing a single namespace could miss mapper-transformed target keys and drop a layer to BF16. Now one shared resolver across all lookup sites.

Thanks to @josephduerr for both reports.

Verified before release

Wheel installed non-editably from PyPI into a clean container, kernels JIT-built from dist-packages/gridbook/csrc, entry point registered, and the 27B artifact served from that install with output byte-identical to a source checkout. GPU battery 363 passed / 0 failed.

Known limits

tp=1 only. Blackwell sm_120/sm_121 for the native path; other NVIDIA GPUs fall back to correct-but-slow kernels. See the compatibility table in the README.