Upstream source provenance

Hangry Labs maintains the model implementation and accelerated inference source in this repository. Builds install both local packages from this checkout. They do not install qwen-tts, qwen-tts-hf, or faster-qwen3-tts from another publisher, clone another inference repository, or apply remote patches at startup. Model checkpoints remain separate, pinned Hugging Face assets.

Three maintained upstreams

Upstream	Last incorporated revision	Local responsibility
QwenLM/Qwen3-TTS	022e286	Model/configuration/processor, inference wrapper, speaker encoder and 12 Hz speech codec in src/qwen_tts/
andimarafioti/Qwen3-TTS, branch distribution/qwen-tts-hf	af457d9	Transformers 5 compatibility in the same local model package
andimarafioti/faster-qwen3-tts	1094aab, version 0.5.4	Static caches, CUDA graphs, generation and streaming in src/faster_qwen3_tts/

Exact machine-readable revisions ship in upstreams.toml. They record the last incorporated branch baselines, not a claim that we continually track a moving branch.

The compatibility branch contains the official base plus three commits:

1. 00969da: the runtime patch proposed in Qwen PR #360.
2. a466278: packaging and publication of qwen-tts-hf.
3. af457d9: release artifact download correction.

The runtime patch is already included in our imported source. It is not applied twice. Its RoPE initialization/buffer restoration and Mimi full-attention fixes preserve behavior expected by the checkpoints under Transformers 5. The other two commits provide distribution provenance; their publishing workflow is not part of our product.

Local boundaries

• src/qwen_tts/: locally maintained Qwen inference source. Original file names and imports are retained to make upstream differences easier to review.
• src/faster_qwen3_tts/: acceleration and the Hangry Labs Studio. The Studio remains under standalone_ui/; both packages ship in one Python distribution.
• tests/: Studio, cache, local inference, compatibility and GPU parity checks.
• scripts/: Docker lifecycle, notices, model prefetch and speech smoke tooling.
• benchmarks/: manual product measurements; benchmarks/research/ retains the inherited NVIDIA investigations separately.

Local changes omit the upstream Gradio demo/launcher, training tools and package publisher. The unused 25 Hz codec and its assets/imports are omitted; this Studio supports the five catalogued 12 Hz checkpoints. The 12 Hz tokenizer wrapper is adapted accordingly and the local package version identifies Hangry Labs changes. GGML/GGUF/native platform adapters, host setup scripts and local-ASR helpers are also removed. General libraries such as Torch, Transformers and Librosa remain ordinary dependencies. Gradio, ONNX Runtime, Python SoX and TorchAudio are not required by this runtime.

The streaming fork and other inspiration credited in NOTICE remain part of our source attribution. They are distinct from these three ongoing upstream synchronization sources.