FlashInfer
Copyright 2025-2026 NVIDIA
Copyright 2023-2026 FlashInfer community (https://flashinfer.ai/)

-------------------------------------------------------------------------------------------------
Some of the code in this project are adapted from other open-source projects with different
licenses. This product also bundles some third-party components under other open source licenses.
See licenses/ for text of these licenses.

The persistent CuTe DSL KDA kernels are adapted from humanfia/kda-for-kda-release,
Copyright (c) 2026 KDA Team, under the MIT License (licenses/LICENSE.kda-for-kda).

The static PTX KDA kernels are adapted from NVlabs/kda (260927-kda-for-kda),
Copyright (c) 2026 KDA Team, under the MIT License (LICENSE.kda-for-kda.txt).

The TIRx KDA kernels are adapted from NVlabs/kda (260927-kda-for-kda),
Copyright (c) 2026 KDA Team, under the MIT License (licenses/LICENSE.kda-for-kda).
The split kernel also retains its original Apache-2.0 TIRx-author attribution.
Source: https://github.com/NVlabs/kda/tree/260927-kda-for-kda

The experimental MoK training adapter (BF16 and MXFP8 routed experts) and CUDA
schedules are adapted from cursor/mixture-of-kittens, Copyright 2026 Cursor
Research, under Apache-2.0.
Source: https://github.com/cursor/mixture-of-kittens/tree/caeb2963f855c7ad53bb50c8bbf211086405cb98

The Qwen 3.5 and Qwen 3.8 Flash-Next gated attention blocks (flashinfer/experimental/qwen_3_5_block/_cudnn_frost/,
shared by flashinfer/experimental/qwen_3_8_block/) are a subset of the cuDNN Frontend Python package (python/cudnn
and one file of its test tree), Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES, vendored byte-faithfully
with its import roots rewritten; each file keeps its own SPDX header: 67 files under the MIT License
(licenses/LICENSE.cudnn-frontend-MIT.txt; one of them the file from the test tree), 92 under Apache-2.0 and two under
"Apache-2.0 AND MIT". The pinned
revision, the per-file digests and the declared patches are listed in that directory's VENDOR.md and
VENDOR_MANIFEST.json.
Source: https://github.com/NVIDIA/cudnn-frontend/tree/6b21be0e01ffe408c62399b4381d88639c06ddcd

The SM120 DSA indexer's single-CTA decode selection
(flashinfer/experimental/sm120_dsa_indexer/csrc/decode_select_row.cuh) is
adapted from SGLang's DeepSeek-V4 top-k (TopKRegister and
coarse_bin_lower_bound), Copyright 2023-2024 SGLang Team, under Apache-2.0.
Source: https://github.com/sgl-project/sglang (python/sglang/kernels/jit/include/sgl_kernel/deepseek_v4/topk_impl.cuh)

The experimental SM100/SM103 HC sum-state backend includes code derived from
SGLang's original HC implementation, Copyright 2023-2024 SGLang Team,
under Apache-2.0. The low-M GEMM retains its original NVIDIA 2026 copyright
and FlashInfer PR #4266 provenance. The fixed-order batched GEMM retains
code derived from batch_invariant_ops, Copyright 2025 Thinking Machines Lab,
under the MIT License (licenses/LICENSE.batch-invariant-ops.txt).
Pinned sources and modifications: flashinfer/experimental/hc_sum_sm100/SOURCES.md.
