ReVolta attribution notice
--------------------------

ReVolta is derived from 1CatAI/1Cat-vLLM (1Cat-vLLM), which is derived from
vllm-project/vllm. The Python import name remains "vllm". The upstream engine
code uses Apache License 2.0, subject to the component exceptions below.

Sources:
  https://github.com/1CatAI/1Cat-vLLM
  https://github.com/vllm-project/vllm
Copyright contributors to the vLLM project; 1Cat AI and contributors.
ReVolta modifications are maintained by Leon under Apache License 2.0,
except where an existing component license applies. See LICENSE.

This inventory covers retained source and configured build dependencies.
It does not grant rights to model weights or other installed dependencies.
Per-file copyright, license and modification notices remain authoritative.
LICENSES/ contains verbatim upstream texts and SOURCES.json records their
immutable URLs and SHA-256 digests. "Evidence snapshot" identifies a checked
upstream revision, not a claim that all local adaptations are an exact copy.
Remaining source-history gaps are stated explicitly below.

Included components
-------------------

1. TurboMind / LMDeploy (OpenMMLab)
   Source: https://github.com/InternLM/lmdeploy
   Location: csrc/sm70_turbomind/lmdeploy/src/turbomind/
   Also adapted from TurboMind, as stated in their headers:
   csrc/sm70_turbomind/ops/awq_sm70_gemm.cu (s884h kernel integration) and
   the alternative e2m1 decoder in csrc/sm70_turbomind/ops/
   skinny_moe_qpn_sm70.cu.
   License: Apache-2.0 for LMDeploy-origin code; LICENSES/lmdeploy-dated.txt.
   Exception: the locally added kernels/gemm/lattice_codebooks.h identifies
   MIT-licensed tables from gguf-py 0.19.0 / llama.cpp. Keep its SPDX notice
   and the original MIT text recorded in item 7; it is not Apache-2.0 code.
   Copyright OpenMMLab; Shanghai AI Laboratory; NVIDIA for derived utilities,
   as recorded in individual source headers.
   Evidence snapshot: 9b3b1933ee2ad1731a6e49b99867b52925b2a54b, the latest
   upstream commit before the local import on 2026-06-14. Of 141 retained
   files, 89 match that snapshot byte-for-byte; 46 differ and 6 are absent
   upstream at that revision.
   A second file-by-file Git blob comparison against upstream commit
   3bf75ff9fc745798f518e22c5e9e92a2bbf39d23 finds 112 of the 141 retained
   files byte-identical, 25 different and 4 absent. This closer snapshot is
   additional provenance evidence, not an exact base for modified files.
   Revision gap: local import ee0bc6a875f891ff4785bfbdb1cd9ed35b21f647
   records no upstream pin. Repository history, source headers and the
   dated upstream tree establish provenance but not one exact base for all
   modified files. Preserve the later local adaptations and their notices.

2. v100-skinny / dnv2003
   Location: csrc/sm70_turbomind/ops/
   License: MIT; Copyright (c) 2026 v100-skinny contributors.
   Evidence: csrc/sm70_turbomind/ops/LICENSE.v100-skinny and derivation
   notices in skinny_moe_qpn_sm70.cu and awq_qpn_m1_sm70.cu.
   Apache-2.0 integration headers do not replace this MIT notice.

3. FlashAttention-V100 / D. Skryabin and the separate SM70 FA2 port
   Source: https://github.com/ai-bond/flash-attention-v100
   Locations: flash-attention-v100/ (kernel/h3/ is CUTLASS-derived, see
   item 4); csrc/attention/sm70_grouped_long/.
   csrc/attention/sm70_v37/ and csrc/attention/sm70_79t/ are 1Cat SM70
   extensions compiled into the fetched SM70 FA2 target. Their files carry
   1CatAI or vLLM-contributor notices (Apache-2.0 or BSD-3-Clause) and
   include headers from that FA2 port; their CUTLASS-derived files are
   listed in item 4. No file in these two directories carries a
   D.Skryabin notice.
   License: BSD-3-Clause; Copyright (c) 2025, D.Skryabin.
   Existing component LICENSE files and source notices are retained.
   Upstream evidence snapshot: 0ababf82e3836c1786d7937c6002670ca4ecda62;
   full upstream text: LICENSES/flash-attention-v100.txt.
   Revision gap: import ee0bc6a875f891ff4785bfbdb1cd9ed35b21f647 and later
   adaptation commits do not identify a single original upstream base.
   The dated upstream tree and local file history were checked; the modified
   v37, 79T and grouped kernels are not asserted to equal that snapshot.
   Separately, CMake fetches https://github.com/zhinianqin/flash-attention-v100
   at c2eda5e6115b98c3ba4bfd181570668742eece22. Its BSD-3-Clause license and
   AUTHORS are LICENSES/flash-attention-sm70{,-authors}.txt. Its CUTLASS
   submodule is 62750a2b75c802660e4894434dc55e839f322277; see item 4.
   Preserve the distinct Apache-2.0 integration notices and local patches.

4. NVIDIA CUTLASS
   Source: https://github.com/NVIDIA/cutlass
   License: BSD-3-Clause for the identified adapted headers.
   Main build reference: v4.4.2, da5e086dab31d63815acafdac9a9c5893b1c69e2;
   LICENSES/cutlass.txt also retains that upstream file's additional notices.
   flash-attention-v100/kernel/h3/UPSTREAM.md records this same revision.
   csrc/quantization/machete/machete_mainloop.cuh explicitly identifies
   06b21349bcf6ddf6a1686a47a137ad1446579db9; its original license, including
   NVIDIA's copyright, is LICENSES/cutlass-machete.txt.
   The fetched SM70 and SM75 FA2 ports use CUTLASS submodule
   62750a2b75c802660e4894434dc55e839f322277; LICENSES/cutlass-fa2.txt.
   Per-file lineage (checked 2026-10-19 against CUTLASS 62750a2b75,
   v3.8.0, v3.9.2, v4.0.0, v4.1.0, v4.2.1, v4.3.0 and v4.4.2):
   - csrc/attention/sm70_v37/{gemm_with_softmax.h,
     epilogue_visitor_with_softmax.h,reduce_softmax_final.h} and
     csrc/attention/sm70_79t/include/cutlass/gemm/threadblock/mma_pipelined.h
     are modified copies of examples/35_gemm_softmax/gemm_with_softmax.h,
     include/cutlass/epilogue/threadblock/epilogue_visitor_with_softmax.h,
     include/cutlass/reduction/kernel/reduce_softmax_final.h and
     include/cutlass/gemm/threadblock/mma_pipelined.h. They keep the
     2017-2025 NVIDIA notice. The upstream text is identical in every
     checked 2025 revision, including the FA2 submodule, so the exact
     import revision is not recorded. Text: LICENSES/cutlass-fa2.txt.
   - csrc/quantization/marlin/sm70_marlin_mma.cuh adapts CUTLASS
     MmaPipelined (include/cutlass/gemm/threadblock/mma_pipelined.h).
     The epilogue classes in csrc/quantization/marlin/sm70_marlin_splitk.cuh
     and csrc/moe/marlin_moe_wna16/sm70_marlin_gemm.cuh adapt the warp
     offset and partition-K reduction of
     include/cutlass/epilogue/threadblock/{epilogue_base.h,epilogue.h}.
     Import revision not recorded; these files build against v4.4.2.
     Text: LICENSES/cutlass.txt.
   - flash-attention-v100/kernel/h3/{default_fmha.h,fmha.h,prefetch_mma.h}
     derive from v4.4.2 examples/41_fused_multi_head_attention; see
     kernel/h3/UPSTREAM.md.
   - The SM70 FA2 build patches the submodule's
     examples/35_gemm_softmax/gemm_with_softmax.h
     (cmake/patches/sm70_flash_attn_d256_gqa_arch.patch).
   Keep embedded conditions and disclaimers with each header.
   Removal candidate: Machete's Hopper-only native sources are outside the
   SM70 build. They remain present; deletion would require adapting shared
   CMake, dispatch and tests when picking upstream fixes. CUTLASS itself is
   used by SM70 and is not a removal candidate.

5. Marlin / Elias Frantar; Neural Magic modifications
   Source: https://github.com/IST-DASLab/marlin
   Locations: csrc/quantization/marlin/; csrc/moe/marlin_moe_wna16/.
   License: Apache-2.0; LICENSES/marlin.txt, checked at upstream
   1f25790bdd49fba53106164a24666dade68d7c90.
   Copyright (C) Marlin.2024 Elias Frantar; Modified by Neural Magic.
   Revision gap: the locally retained headers name Marlin but no original
   upstream commit. Local rename history and vLLM PR #14447 (merge
   d06ba4ed3f9a5929eabb404842a5c02da42e960b) identify the modified vLLM
   integration, not a single original Marlin revision.
   The sm70_* files in these directories are 1Cat SM70 implementations on
   CUTLASS 2.x APIs. Only sm70_marlin_dispatch.cu and
   sm70_marlin_moe_dispatch.cu carry the Marlin notice;
   sm70_{gptq,awq}_marlin_repack.cu are copies of vLLM's repack kernels.
   Their CUTLASS-derived parts are listed in item 4.

6. NVIDIA FasterTransformer and TensorRT-LLM adaptations
   Sources: https://github.com/NVIDIA/FasterTransformer
            https://github.com/NVIDIA/TensorRT-LLM
   License: Apache-2.0 for the identified files; see LICENSE and their headers.
   csrc/attention/attention_generic.cuh and dtype_*.cuh retain source URLs
   and NVIDIA/vLLM copyright notices. csrc/moe/topk_softmax_kernels.cu
   identifies TensorRT-LLM v0.7.1. Other MoE files retain their own source
   references and notices.

7. llama.cpp / ggml authors and vLLM GGUF plugin
   Sources: https://github.com/ggml-org/llama.cpp
            https://github.com/vllm-project/vllm-gguf-plugin/pull/141
   Licenses: llama.cpp MIT; plugin bridge Apache-2.0.
   Existing texts and provenance: vllm/third_party/gguf_native/ and
   csrc/quantization/gguf_upstream/llama.cpp/LICENSE, source_manifest.json;
   csrc/libtorch_stable/quantization/gguf_upstream/LICENSE.plugin.
   Recorded revisions: llama.cpp 002a12ad25503a93501b2e188c360029830a241a;
   bridge aa09d6522f29325d64d999d7d7c794f79836de07.
   Older csrc/quantization/gguf/ headers identify llama.cpp b2899.

8. FlashQLA / Qwen, Alibaba
   Source: https://github.com/QwenLM/FlashQLA
   Location: flash_qla/
   License: MIT; Copyright (c) 2026 Qwen, Alibaba.
   Texts: flash_qla/LICENSE; LICENSES/flashqla-dated.txt.
   Evidence snapshot: 5be4204363d1f751ba586b0972eab58b818c852c, preceding the
   local import 25733d9cf98d47108ff4b8da2138aa121ad1d1b1 on 2026-07-19.
   Four retained files match; 15 differ and 4 are not in that snapshot.
   Revision gap: the import message and headers contain no exact source pin;
   this is a locally adapted subtree, not an exact upstream checkout.

9. NVIDIA Management Library Python bindings (nvidia-ml-py)
   Source: https://pypi.org/project/nvidia-ml-py
   Location: vllm/third_party/pynvml.py
   License: BSD-3-Clause notice embedded in the file alongside Apache-2.0
   integration headers. Copyright (c) 2011-2023 NVIDIA Corporation.
   The file records version 12.570.86. Retain its original full notice.

10. SGLang CPU kernels
    Source: https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
    Location: csrc/cpu/sgl-kernels/
    License: Apache-2.0 from the component's own sgl-kernel/LICENSE;
    Copyright 2023-2024 SGLang Team.
    Evidence snapshot: a9a8b20a90df3c8b64f7308abde7d3d1fa1a5d12.
    Texts: LICENSES/sglang-kernel.txt and sglang-notices.txt (the component's
    complete upstream THIRDPARTYNOTICES, retained conservatively).
    Revision gap: local update b3945cc3160119111927816841950715c5410708
    (vLLM PR #41924) and its linked source URLs do not specify the original
    SGLang commit. All 13 comparable files differ; 2 have no same-path
    upstream file. The dated tree, import PR and local history were checked.
    Removal candidate: CPU-only sources are outside the SM70 build target.
    They remain present. Deletion would require adapting CPU CMake lists,
    CPU op registration and tests, and would increase conflicts when picking
    upstream CPU or shared build fixes.

11. Flash Linear Attention (FLA) adaptations
    Source: https://github.com/fla-org/flash-linear-attention
    Location: vllm/model_executor/layers/fla/ops/
    License: MIT for copied FLA code; LICENSES/fla-dated.txt, checked at
    8a91ac503764fed5bb65ad6d9b157880eeff2820 before the vLLM import.
    Preserve the additional per-file copyrights: (c) 2023-2025 Songlin Yang,
    Yu Zhang; (c) 2024 Tri Dao for layernorm_guard.py. Apache-2.0 integration
    headers do not replace the copied MIT attribution.
    Revision gap: vLLM PR #24518 and merge
    1aa427fdc1d4ddb7a0c966f7c46badd1c3952c14 name FLA but give no exact
    original revision. The PR, source headers and local import were checked.
    These runtime kernels are not a CPU-only removal candidate.

12. Mamba and causal-conv1d adaptations
    Sources: https://github.com/state-spaces/mamba
             https://github.com/Dao-AILab/causal-conv1d
    Location: vllm/model_executor/layers/mamba/ops/
    Mamba license: Apache-2.0; LICENSES/mamba.txt. Copyright (c) 2023-2024
    Tri Dao, Albert Gu. Source headers cite v2.2.4, resolved to
    95d8aba8a8c75aedcaa6143713b11e745e7cd0d9. layernorm_gated.py cites
    60dadf2e0ee730ac337035d5533de10bc26e4847; its license text is identical.
    causal-conv1d license: BSD-3-Clause; Copyright (c) 2022, the respective
    contributors in AUTHORS; copied adaptation credits (c) 2024 Tri Dao.
    Full texts: LICENSES/causal-conv1d{,-authors}.txt, from v1.5.0,
    fe772c75e29da8123d4478b670342111ce23cc6b.
    Revision gap: causal_conv1d.py links upstream "main"; local history and
    that URL do not establish its exact import revision. v1.5.0 is license
    evidence, not a claimed exact copy. Preserve per-file integration notices.

13. SGLang-V100 SM70 one-shot push all-reduce
    Source: https://github.com/haohervchb/sglang-V100 (a fork of
    https://github.com/sgl-project/sglang), revision 845b9fdf7a7e.
    Location: the SM70 TP4 one-shot push collective in
    csrc/custom_all_reduce.cuh, as stated in its comment.
    License: Apache-2.0 (the fork's repository license, checked
    2026-10-19); text: LICENSE.

Retained but not built for the SM70 (7.0) release
-------------------------------------------------

Machete (csrc/quantization/machete/, 9.0a only); CUTLASS scaled_mm c2x/c3x
and grouped MoE (csrc/libtorch_stable/quantization/w8a8/cutlass/, except the
scaled_mm_entry.cu dispatcher); W4A8
(csrc/libtorch_stable/quantization/cutlass_w4a8/); SM100 CUTLASS MLA
(csrc/libtorch_stable/attention/mla/); NVFP4/MXFP4 kernels
(csrc/libtorch_stable/quantization/fp4/, except nvfp4_quant_entry.cu and
nvfp4_scaled_mm_entry.cu); MXFP8 MoE (csrc/moe/mxfp8_moe/); DSV3 router and
fused-A GEMM; CUTLASS epilogue extensions (csrc/cutlass_extensions/epilogue/,
csrc/libtorch_stable/cutlass_extensions/); TurboMind SM80/SM90 GEMM headers
(csrc/sm70_turbomind/lmdeploy/src/turbomind/kernels/gemm/ *sm90*,
mainloop_sm80_v2.h, scheduler.cuh); non-SM70 Marlin kernels. Their per-file
notices still apply to source distributions.

Build-fetched components
------------------------

For the 7.0 release build, DeepGEMM, FlashMLA, QuTLASS, vLLM FlashAttention
and SM75 FlashAttention are not fetched; CUDA builds fetch CUTLASS v4.4.2,
the SM70 FA2 port (with its CUTLASS submodule) and Triton kernels.

Configured default revisions are listed below. A source-directory override
or a changed pin requires a fresh license and transitive-notice check.
License texts do not establish which optional components a wheel contains.

- vLLM FlashAttention: https://github.com/vllm-project/flash-attention
  bce29425653ec0fbc579d329883030e832d15ada; BSD-3-Clause;
  LICENSES/vllm-flash-attention{,-authors}.txt.
- SM75 FlashAttention: https://github.com/Peuqui/flash-attention
  43b9d29c9aa8e18d9351e7c643dc78ef7e7979fc; BSD-3-Clause;
  LICENSES/flash-attention-sm75{,-authors}.txt.
- FlashMLA: https://github.com/vllm-project/FlashMLA
  a6ec2ba7bd0a7dff98b3f4d3e6b52b159c48d78b; MIT; LICENSES/flashmla.txt.
- DeepGEMM: https://github.com/deepseek-ai/DeepGEMM
  891d57b4db1071624b5c8fa0d1e51cb317fa709f; MIT; LICENSES/deepgemm.txt.
- Triton kernels: https://github.com/triton-lang/triton
  v3.5.1, 0add68262ab0a2e33b84524346cb27cbb2787356; MIT; LICENSES/triton.txt.
- QuTLASS: https://github.com/IST-DASLab/qutlass
  830d2c4537c7396e14a02a46fbddd18b5d107c65; Apache-2.0;
  LICENSES/qutlass.txt.

The referenced CMake files and upstream LICENSE/AUTHORS at these pins were
checked. Their top-level license texts are now included. Distribution gap:
this source inventory is not an audit of an assembled wheel or container's
complete transitive dependency contents and notices. Recheck the exact
release artifacts, optional components and submodules before redistribution.
