feat: add OmniVoice TTS model support#147

Draft

leegang wants to merge 113 commits intoBlaizzy:mainfrom

leegang:feature/omnivoice-support

Commits on Apr 11, 2026

chore: simplify CLI by removing OmniVoice-specific parameters

bestmrlee-max
and
qwencoder
committed
refactor: replace setGenerationConfig with generate() overload

bestmrlee-max
and
qwencoder
committed
Merge branch 'Blaizzy:main' into feature/omnivoice-support
leegang
authored

Commits on Apr 12, 2026

fix: resolve audio_embeddings shape mismatch for OmniVoice checkpoint

bestmrlee-max
and
qwencoder
committed
fix: resolve audio_heads shape mismatch for OmniVoice checkpoint

bestmrlee-max
and
qwencoder
committed
fix: add backbone key to llm ModuleInfo for checkpoint compatibility

bestmrlee-max
and
qwencoder
committed
fix: map backbone. weight prefix to model. for OmniVoice

bestmrlee-max
and
qwencoder
committed
fix: correct weight key mapping and concatenate per-codebook audio embeddings

bestmrlee-max
and
qwencoder
committed
fix: stop concatenating per-codebook audio embeddings/heads

bestmrlee-max
and
qwencoder
committed
fix: map unprefixed LLM weights to llm.model. path

bestmrlee-max
and
qwencoder
committed
fix: expose Qwen3ModelInner via @ModuleInfo for weight loading

bestmrlee-max
and
qwencoder
committed
fix: match @ModuleInfo key to sanitized weight prefix

bestmrlee-max
and
qwencoder
committed
fix: skip audio tokenizer weight loading (TODO)

bestmrlee-max
and
qwencoder
committed
feat: implement full HiggsAudioV2 audio tokenizer with DAC encoder/decoder and RVQ

bestmrlee-max
and
qwencoder
committed
fix: match OmniVoiceAudioTokenizer quantizer key to checkpoint ('quantizers')

bestmrlee-max
and
qwencoder
committed
fix: align OmniVoice encoder/decoder blocks with Higgs Audio V2 architecture

bestmrlee-max
and
qwencoder
committed
fix: add missing encoder tail layers (snake1, conv1, conv2) for OmniVoice acoustic encoder

bestmrlee-max
and
qwencoder
committed
fix: use PyTorch-weight-layout ConvTranspose1d for OmniVoice decoder

bestmrlee-max
and
qwencoder
committed
fix: add missing decoder tail layers (snake1, conv1, conv2) for OmniVoice acoustic decoder

bestmrlee-max
and
qwencoder
committed
fix: completely restructure encoder/decoder/quantizer to match Higgs Audio V2 checkpoint

bestmrlee-max
and
qwencoder
committed
fix: use OmniVoiceConv1d with PyTorch weight layout [out, in, kernel] throughout

bestmrlee-max
and
qwencoder
committed
fix: bias reshape to (1,1,-1) for NLC tensor layout

bestmrlee-max
and
qwencoder
committed
fix: use reshaped([-1]) for bias to ensure correct broadcasting with NLC tensors

bestmrlee-max
and
qwencoder
committed
fix: reshape bias to (1,1,nChannels) for explicit NLC broadcasting

bestmrlee-max
and
qwencoder
committed

Commits on Apr 13, 2026

Commits on Apr 14, 2026

Commits on Apr 15, 2026