v0.0.14 2.0.3
droidrun/mobilerunMay 11, 2026 → May 11, 20260 days between releases
What's New
Major version jump to 2.0.3 adds fine-tuning validation, ref_audio support, explicit device selection, and faster streaming VAE decoding.
Breaking Changes
- The jump from 0.0.14 to 2.0.3 is a major version change and likely carries incompatible API/behavior changes not fully enumerated in the notes.
- Legacy LoRA checkpoints are now loaded with weights_only=True, so checkpoints requiring arbitrary code execution will fail to load.
- Training manifest packing now supports an optional ref_audio segment, which may alter expected input layout for existing pipelines.
New Features
- Added `voxcpm validate` to pre-flight check JSONL training manifests for format, required fields, audio readability, sample rate, and duration/text stats.
- Added optional `ref_audio` support in the fine-tuning data pipeline, with loss applied only to the target audio segment.
- Added a `--device` CLI argument supporting auto, cpu, mps, cuda, and indexed CUDA devices for model inference commands.
- Introduced a stateful StreamingVAEDecoder for VoxCPM2 streaming inference.
Improvements
- Streaming VAE decoding now processes only the newest latent patch and carries causal convolution state, removing redundant overlap decoding.
- Runtime device handling improved with safer MPS dtype behavior.
- LoRA checkpoint loading hardened via weights_only=True for safer deserialization.
Bug Fixes
- Fixed LoRA rank mismatch handling in lora_ft_webui.py.
- Corrected MPS dtype behavior during inference.
Upgrade Notes
Treat this as a major upgrade: re-validate existing training manifests with `voxcpm validate` and confirm legacy LoRA checkpoints still load under weights_only=True. Update inference scripts to pass `--device` explicitly, especially on Apple MPS or multi-GPU hosts. The 0.0.x publishing/quantization features from v0.0.14 are not mentioned in 2.0.3, so verify they remain available before upgrading.
The Two Releases
v0.0.14
May 11, 2026· 5 months ago
Introduces publishing capabilities for wk-st fine-tunes to HuggingFace and optimizes int8 quantization to reduce model size.
Full release notesv2.0.3: fine-tuning validation, runtime stability, and streaming improvements
May 11, 2026· 5 months ago
Focuses on fine-tuning usability, runtime stability, and streaming improvements for VoxCPM2.
Full release notes