THIRD_PARTY_NOTICES.md

NEO mixed-GGUF native runtime notices

Paiton runtime artifacts and the public plugin are distributed under the root Apache-2.0 license. The proprietary compiler and generated implementation source are not distributed. The rights-holder authorization for generated runtime binaries recorded in the existing Qwen3.8 release process applies.

The GGUF format and quantization algorithms are based on ggml/llama.cpp revision 434ddbbc0e30522e897670681e503b797c12b7c1. Preserve the ggml authors' MIT notice in LICENSES/ggml-MIT.txt. The compiled attention path uses AMD rocWMMA from the pinned ROCm 7.14 build; its MIT notice is included in LICENSES/rocWMMA-MIT.txt. Existing root LICENSES retain notices for inherited runtime dependencies, including Composable Kernel, AITER, flash-linear-attention and Triton. These inherited packages belong to the external vLLM stack; the new Paiton model is compiled from native HIP/C++ without Torch or Triton imports or generated Triton kernels.

The native vision encoder follows the Qwen3-VL/Qwen3.5 architecture and the selected GGUF projector contract. Its HIP positional interpolation, layer normalization, rotary attention and merger replace the external vision model; vLLM and Transformers provide image preprocessing and API integration under their existing Apache-2.0 notices. No Triton vision kernel is compiled or embedded into the Paiton artifact.

The runtime reuses the pinned public vLLM/ROCm image documented in the Dockerfile. Its notices remain under /licenses. The NEO image removes the unrelated Qronos payload and Transformers' test-only testing_utils fixture, which embeds a sandbox CI token. No production model weights or tokenizer values are altered by this cleanup. Runtime dependency licenses also remain with their installed packages; the release's dependency SBOM records their exact versions.

Model weights are downloaded separately from the author's pinned repository. The GGUF and author-source repositories declare Apache-2.0. Exact repository revisions, file hashes, tokenizer provenance and the selected mixed tensor formats are recorded in checkpoint.lock.json. Paiton preserves the selected GGUF packed weight values; it does not claim authorship of the fine-tune or redistribute it as a new quantization. Neither the base Qwen checkpoint nor the AMD Qronos fine-tune substitutes for these weights. MTP tensors remain in the upstream file but are not loaded into the qualified target-only model.