Accurate, fast FP4 KV-cache decoding through algorithm-system co-design.
Third-party notices
===================
The MIT license in LICENSE applies to original BitDecoding-FP4 code.
Third-party components retain their own copyright notices and licenses.
FlashAttention-derived helpers
------------------------------
The helpers under src/nvfp4_decode_kernel/_fa4/ include code derived from
FlashAttention, except for the NVIDIA helper identified below.
Upstream: https://github.com/Dao-AILab/flash-attention
License: BSD 3-Clause; see src/nvfp4_decode_kernel/_fa4/LICENSE.
Upstream authors: see src/nvfp4_decode_kernel/_fa4/AUTHORS and the copyright
notices retained in the individual source files.
NVIDIA CUTLASS DSL helper
------------------------
File: src/nvfp4_decode_kernel/_fa4/block_scaled_layout.py
Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
This software contains source code provided by NVIDIA Corporation.
This file is governed by the NVIDIA CUTLASS DSL Software License Agreement,
not the MIT or BSD license grants:
https://docs.nvidia.com/cutlass/latest/media/docs/pythonDSL/license.html
vLLM submodule
--------------
The third_party/vllm submodule is licensed under Apache License 2.0.
Its license is included at third_party/vllm/LICENSE when the submodule is
checked out. Third-party components within vLLM retain their own licenses.
Upstream: https://github.com/vllm-project/vllm