Third-party notices
This index identifies existing notices for bundled source components. The referenced texts and file headers remain authoritative for their respective components. This is not a complete audit of Cargo dependencies, Python packages, CUDA libraries, model weights, or container base images.
| Component | Retained license and attribution locations |
|---|---|
| fastsafetensors non-GDS reader | Apache-2.0; license and pinned source/adaptation provenance. Only the native worker reader is bundled; no Python, GDS, libaio, io_uring or NUMA dependency. |
| Hugging Face Text Embeddings Inference | LICENSE, NOTICE, and inherited file headers. |
| Vendored TEI regression-test snapshots | Apache-2.0 comparison baselines; Candle snapshot provenance and router snapshot provenance list the fixtures, full Git blob hashes, and pinned upstream revisions. |
| Adapted Candle CPU operators | Apache text and source attribution in ModernBERT CPU math, based on Candle revision 036f39e. |
| Candle CUDA activation kernel | MIT text and kernel notices. |
| Native prefix Q/K normalization and rotary kernel | Derived from Candle Q/K normalization, with the MIT permission notice retained in the derived kernel. |
| Candle cuBLASLt extension | Apache text and MIT text. |
| Candle index-select extension | Apache text, MIT text, and file headers. |
| Candle layer-norm extension | Apache text, MIT text, and BSD text. |
| Candle rotary extension | Apache text and MIT text. |
| RadixMLP-derived layers | MIT attribution in index-select and RadixMLP source headers. |
| CUTLASS-derived Gemma4 and Nomic MoE kernels | BSD-3-Clause text. |
| xxHash prefix fingerprints | BSD-2-Clause text, upstream header notices, and pinned source provenance. |
| ALiBi implementation | Source notices for Google AI Language Team, Hugging Face, and Jina AI. |
The native Torch embedding normalization kernel
derives from the bundled Candle layer-norm extension's
kernels/ln_fwd_kernels.cuh. It retains the BSD-3-Clause copyright and license
header; the extension's BSD license text is also installed with the native library.
Basetenkenizer is a Cargo dependency rather than vendored source. Its published package carries Apache-2.0 licensing and additional component notices; consult its license audit for the version used by the release. Other dependencies must likewise be audited from their actual pinned versions when assembling distributed artifacts.
The native Nomic eight-expert router in backends/libtorch/cpp/kernels/fast_kernels.cu
derives its strict bitonic ordering from Candle sort.cu at revision
036f39e44015e78ba18d716d73dfdc2e78730c44, adapted upstream from llama.cpp.
The selected MIT permission notice is retained beside the derived kernel.
Mistral3/Pixtral model equations and processor behavior are adapted from
Hugging Face Transformers 5.12.1 (Apache-2.0). The retained license and source
attribution are in backends/libtorch/cpp/vendor/transformers_pixtral/.
The packed CUDA Gated DeltaNet chunkwise transform is adapted from
Hugging Face Transformers 5.19.0 (Apache-2.0). The retained license, source hash,
and adaptation notice are in backends/libtorch/cpp/vendor/transformers_gdn/.
Hopper Gated DeltaNet prefill uses FlashInfer 0.6.13 at revision
57ba7eeb7ea3003a2d6ad5d9a057c4f952709bac (Apache-2.0). Its LICENSE, NOTICE,
and third-party license texts are installed under share/licenses/tei_torch/flashinfer/.
Its device launcher is locally patched to initialize shared-memory opt-in on each CUDA device/context.