Add Qwen3 TTS architecture support#20752
Open
Acceldium wants to merge 3 commits intoggml-org:masterfrom
Open
Conversation
Add GGUF support for the Qwen3-TTS model family: - gguf-py: define QWEN3TTS and QWEN3TTS_CP architectures, TTS-specific KV keys (text_vocab_size, text_embedding_length, num_code_groups, position_id_per_seconds), and tensor name mappings - convert_hf_to_gguf.py: add Qwen3TTSTalkerModel (28-layer Talker with interleaved MRoPE) and Qwen3TTSCodePredictorModel (5-layer Code Predictor with standard RoPE), including speaker encoder tensor remapping - tools/tts/convert_qwen3tts.py: wrapper script to convert both Talker and Code Predictor GGUFs from a single HF model directory - tools/tts/convert_qwen3tts_tokenizer.py: tokenizer conversion helper
Add C++ inference support for Qwen3-TTS Talker and Code Predictor: - llama-arch: register LLM_ARCH_QWEN3TTS (Talker, 28-layer with MRoPE) and LLM_ARCH_QWEN3TTS_CP (Code Predictor, 5-layer with standard RoPE), including KV keys, tensor names, and tensor info entries - llama-hparams: add TTS-specific fields (text_vocab_size, text_embd_size, num_code_groups, position_id_per_s) - llama-model: implement load_hparams, load_tensors for both architectures with TTS tensor pointers (text/codec embeddings, projection layers, per-codebook heads) - llama-graph: guard build_inp_embd for null tok_embd to support embedding-only input path used by TTS - models/qwen3tts.cpp: llm_build_qwen3tts graph builder with text projection, codec embedding, and MRoPE attention - models/qwen3tts_cp.cpp: llm_build_qwen3tts_cp graph builder for the code predictor sub-model
Add the end-user facing components for Qwen3-TTS: - tools/tts/qwen3tts.cpp: main CLI tool implementing the full TTS pipeline -- text tokenization, Talker prefill/decode with MRoPE, Code Predictor for multi-codebook generation, and vocoder (DAC-based decoder with VQ, strided ConvTranspose1d, Snake activations) - tools/tts/speaker-encoder: standalone ECAPA-TDNN speaker encoder for voice cloning (extracts speaker embeddings from reference audio) - tools/tts/CMakeLists.txt: build targets for llama-qwen3tts and speaker-encoder - examples/qwen3-tts/: Python usage examples (basic TTS, voice cloning, multilingual, Gradio app, benchmarking)
|
Hi @Acceldium, thanks for your contribution! Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:
Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below. |
github-actions
bot
added
model
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Note that this branch adds initial support for running Qwen3-TTS models in llama.cpp. Qwen3-TTS uses a multi-stage pipeline (language model + audio decoder/tokenizer) that requires executing multiple independent compute graphs in sequence — a pattern llama.cpp does not currently support natively.