CITATION.cff

cff-version: 1.2.0
message: "If you use gridbook or the NVFP4-CB / FP8-CB formats, please cite this repository."
title: "gridbook: product-codebook weight formats that serve on native tensor-core kernels"
abstract: >-
  gridbook specifies the NVFP4-CB and FP8-CB product vector-quantization weight
  formats and provides an out-of-tree vLLM plugin that serves them. Codebook
  entries lie exactly on a hardware grid (NVFP4 or FP8), so a decoded tile is
  bit-compatible with a native format and is consumed by the same tensor-core
  GEMM path a conventional checkpoint uses, without a forked runtime.
type: software
authors:
  - family-names: Tand
    given-names: Robert
    email: robert.tand@icloud.com
repository-code: "https://github.com/RobTand/gridbook"
url: "https://github.com/RobTand/gridbook"
license: Apache-2.0
version: "0.9.1"
date-released: 2026-08-30
keywords:
  - quantization
  - vector-quantization
  - codebook
  - llm-inference
  - vllm
  - cuda
  - nvfp4
  - fp8
  - blackwell