Skip to content

LICENSING.md

Latest commit

 

History

History
176 lines (155 loc) · 21.9 KB

LICENSING.md

File metadata and controls

176 lines (155 loc) · 21.9 KB
Raw

Licensing

cudnn-frontend is distributed primarily under the Apache License 2.0 (see LICENSE.txt). Some files retain MIT or BSD-3-Clause terms for pre-existing code. Every source file carries an SPDX SPDX-License-Identifier: expression declaring the applicable licenses.

To find the license of any file, read its SPDX tag, e.g.:

SPDX-License-Identifier: Apache-2.0
SPDX-License-Identifier: MIT
SPDX-License-Identifier: Apache-2.0 AND MIT
SPDX-License-Identifier: Apache-2.0 AND BSD-3-Clause

Third-party attributions are in THIRD_PARTY_LICENSES.txt.

Apache-2.0 modifications to existing licensed code

The OSS warning cleanup licenses its NVIDIA modifications under Apache-2.0, including changes to files with retained MIT or BSD-3-Clause code. Their headers identify the scope of the modifications and use an SPDX AND expression to preserve the pre-existing terms. This is not an alternative-license choice: both sets of applicable terms must be respected. See the SPDX license-expression specification.

Original copyright notices, provenance comments, and embedded license texts remain intact. MIT terms are reproduced in LICENSE-MIT.txt, and upstream license texts and attributions are reproduced in THIRD_PARTY_LICENSES.txt. Where an existing file carries both an MIT declaration and a BSD notice, its expression retains both. The adapted CUTLASS static scheduler in python/cudnn/_cutlass_helpers/ likewise retains BSD-3-Clause for the upstream code and applies Apache-2.0 to local changes.

An Apache-2.0 modification does not assert that upstream authors relicensed their contributions. Files whose complete NVIDIA ownership is established can use Apache-2.0 alone; _fp8_quant.py in block_sparse_attention/ is one such file.

Why some files remain under MIT

The repository's original license was MIT. During the MIT → Apache 2.0 relicensing (PR #408, sign-off issue #431), files were converted to Apache 2.0 only when all of their surviving code is owned by NVIDIA. Files are kept under MIT in two cases:

  1. Pending external-contributor consent — the file contains code contributed by a non-NVIDIA contributor whose consent to relicense has not (yet) been obtained. Determined by git blame on develop: a file stays MIT if any external contributor's lines survive in it.
  2. Third-party-derived code — the file is derived from external open source (FlashAttention, QuACK, or CUTLASS) and carries the original author's copyright.

If/when a listed external contributor grants consent, the files attributed to them below can be moved to Apache 2.0 by flipping their SPDX tag.

Consent received so far (see issue #431): @take-cheeze, @fallintoplace, @zianglih, @JackRao123, @zkyue, @Hyaloid, @haowen-han, @junaire, @szluyu99, @dimitar-asenov — their files have already been moved to Apache-2.0. A file still appears below if another contributor who has not yet consented also has surviving lines in it.

Cleared by NVIDIA employment (not by issue #431 consent): @HollowMan6, @hxbai — commits under a personal email address but is an NVIDIA employee, so those contributions are covered by employment and their files are Apache-2.0.

The Introducing commit(s) column links the exact commit that introduced the surviving external line(s) in each file (blame on origin/develop).

Category 1 — MIT pending external-contributor consent (51 files)

File External contributor(s) Introducing commit(s)
CMakeLists.txt Connor Baker (@ConnorBaker) 0f828cf (#125)
cudnn_frontend-config.cmake.in Connor Baker (@ConnorBaker) 0f828cf (#125)
dlpack_version.txt Emilien Macchi (@EmilienM) 1669048 (#165)
include/cudnn_frontend/graph_properties.h Subhobrata Dey (@sbcd90) 6943af9 (#214)
include/cudnn_frontend/node/conv_dgrad.h DrDirk (@DrDirk) 5f680bc (#423)
include/cudnn_frontend/node/conv_fprop.h DrDirk (@DrDirk) 5f680bc (#423)
include/cudnn_frontend/node/conv_wgrad.h DrDirk (@DrDirk) 5f680bc (#423)
include/cudnn_frontend/node/matmul.h James Y Knight (@jyknight) af9bc9e (#56)
include/cudnn_frontend/node/pointwise.h DrDirk (@DrDirk)
James Y Knight (@jyknight)		 5f680bc (#423)
af9bc9e (#56)
include/cudnn_frontend/node/reduction.h James Y Knight (@jyknight) af9bc9e (#56)
include/cudnn_frontend/node/reshape.h James Y Knight (@jyknight) af9bc9e (#56)
include/cudnn_frontend/node/rng.h James Y Knight (@jyknight) af9bc9e (#56)
include/cudnn_frontend/node/softmax.h James Y Knight (@jyknight) af9bc9e (#56)
include/cudnn_frontend/node_interface.h James Y Knight (@jyknight) af9bc9e (#56)
include/cudnn_frontend_utils.h Martin Valgur (@valgur) 31b2c5d (#154)
python/CMakeLists.txt Connor Baker (@ConnorBaker)
Emilien Macchi (@EmilienM)		 0f828cf (#125)
1669048 (#165)
samples/cpp/CMakeLists.txt Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/convolution/dgrads.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/convolution/fp8_fprop.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/convolution/fprop.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/convolution/int8_fprop.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/convolution/wgrads.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/matmul/fp8_matmul.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/matmul/int8_matmul.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/matmul/matmuls.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/matmul/mixed_matmul.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/misc/pointwise.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/misc/resample.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/misc/serialization.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/misc/slice.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/misc/sm_carveout.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/norm/batchnorm.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/norm/layernorm.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/norm/rmsnorm.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp16_bwd.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp16_bwd_with_cudagraphs.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp16_bwd_with_flexible_graphs.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp16_cached.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp16_fwd.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp16_fwd_paged_decode_and_prefill.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp16_fwd_with_cudagraphs.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp16_fwd_with_custom_dropout.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp16_fwd_with_flexible_graphs.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp16_fwd_with_paged_caches.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp8_bwd.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp8_bwd_bottom_right_causal_mask.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp8_fwd.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/cpp/sdpa/fp8_fwd_bottom_right_causal_mask.cpp Connor Baker (@ConnorBaker) 0f828cf (#125)
samples/legacy_samples/CMakeLists.txt Connor Baker (@ConnorBaker) 0f828cf (#125)
setup.py Emilien Macchi (@EmilienM) e0449f6 (#163)
test/cpp/CMakeLists.txt Connor Baker (@ConnorBaker) 0f828cf (#125)

Note: dlpack_version.txt is a plain version-string file that cannot carry a header comment; it is listed here and governed by MIT via this manifest.

Category 2 — Third-party-derived files retaining MIT terms (36 files)

Derived from FlashAttention or CUTLASS (both BSD-3-Clause) and/or QuACK (Apache-2.0); they retain their original authors' copyright notices. See THIRD_PARTY_LICENSES.txt. Files with Apache-2.0 modifications retain MIT in their compound SPDX expression. The commit link(s) are the NVIDIA import commits that introduced the surviving derived lines. This historical list does not enumerate every later import; the source headers and third-party notices also cover those.

File Import commit(s)
python/cudnn/block_sparse_attention/_interface.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/bwd/bsa_bwd_postprocess.py afd575d (#333)
5b9c94a (#349)
python/cudnn/block_sparse_attention/csrc/bwd/bsa_bwd_preprocess.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/bwd/sm100_blk128/bsa_bwd_sm100.py afd575d (#333)
0041a2b (#382)
d380fab (#350)
python/cudnn/block_sparse_attention/csrc/fwd/sm100_blk64/bsa_fwd_combine.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/fwd/sm100_blk64/bsa_fwd_helpers.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/utils/block_info.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/utils/block_sparse_tile_scheduler.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/utils/copy_utils.py afd575d (#333)
0041a2b (#382)
python/cudnn/block_sparse_attention/csrc/utils/cute_dsl_utils.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/utils/kernel_utils.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/utils/layout_utils.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/utils/mma_sm100_desc.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/utils/named_barrier.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/utils/pack_gqa.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/utils/pipeline.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/utils/sm90_utils.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/utils/softmax.py afd575d (#333)
52119ee (#341)
python/cudnn/block_sparse_attention/csrc/utils/tcgen05_mma_helpers.py afd575d (#333)
python/cudnn/block_sparse_attention/csrc/utils/tile_scheduler.py afd575d (#333)
python/cudnn/deepseek_sparse_attention/score_recompute/dense_score_recompute_sm90.py c4a9762 (#241)
3e2b8ec (#263)
7016b04 (#316)
28462c3 (#273)
python/cudnn/deepseek_sparse_attention/score_recompute/sparse_score_recompute_sm90.py c4a9762 (#241)
python/cudnn/deepseek_sparse_attention/sparse_attention_backward/_interface_sm100.py c4a9762 (#241)
7016b04 (#316)
3e2b8ec (#263)
cfee724 (#318)
python/cudnn/deepseek_sparse_attention/sparse_attention_backward/_interface_sm90.py c4a9762 (#241)
3e2b8ec (#263)
f3ee97b (#388)
python/cudnn/deepseek_sparse_attention/utils/copy.py c4a9762 (#241)
python/cudnn/deepseek_sparse_attention/utils/sm90/bwd_barriers.py c4a9762 (#241)
python/cudnn/deepseek_sparse_attention/utils/sm90/bwd_tile_scheduler.py c4a9762 (#241)
python/cudnn/deepseek_sparse_attention/utils/sm90/mma.py c4a9762 (#241)
74efc0d (#321)
python/cudnn/deepseek_sparse_attention/utils/sm90/primitives.py c4a9762 (#241)
7016b04 (#316)
python/cudnn/hstu/hstu_attention/_kernels/blackwell_helpers.py c5471cb
python/cudnn/hstu/hstu_attention/_kernels/block_info.py c5471cb
python/cudnn/hstu/hstu_attention/_kernels/fast_math.py c5471cb
python/cudnn/hstu/hstu_attention/_kernels/mask.py c5471cb
python/cudnn/hstu/hstu_attention/_kernels/mma_sm100_desc.py c5471cb
python/cudnn/hstu/hstu_attention/_kernels/tile_scheduler.py c5471cb
python/cudnn/hstu/hstu_attention/_kernels/utils.py c5471cb