Accurate, fast FP4 KV-cache decoding through algorithm-system co-design.
[build-system]
requires = ["setuptools>=77"]
build-backend = "setuptools.build_meta"
[project]
name = "bitdecoding-fp4-kernel"
version = "0.1.0"
description = "NVFP4 paged-decode kernel interface"
license = "MIT AND BSD-3-Clause AND LicenseRef-NvidiaProprietary"
license-files = [
"LICENSE",
"NOTICE",
"src/nvfp4_decode_kernel/_fa4/LICENSE",
"src/nvfp4_decode_kernel/_fa4/AUTHORS",
]
requires-python = ">=3.10"
dependencies = [
"torch",
"nvidia-cutlass-dsl==4.6.0",
"quack-kernels==0.6.1",
"cuda-python",
"typing_extensions",
"apache-tvm-ffi==0.1.10",
"torch-c-dlpack-ext",
]
# Points `backend: CUSTOM` at NVFP4Backend in every vLLM process.
[project.entry-points."vllm.general_plugins"]
nvfp4_vllm = "nvfp4_vllm:register"
[project.optional-dependencies]
eval = [
"flashinfer-python",
"datasets",
"pandas",
"pyarrow",
"safetensors",
# RULER's synthetic-data generator.
"nltk",
]
[tool.setuptools.packages.find]
where = ["src"]