LoRA is a first-class workflow in VoxCPM, and the project already prefers
safetensors plus weights-only fallback loading for base model artifacts. The
legacy LoRA .ckpt/.pth path was the remaining place that still deserialized
arbitrary pickle objects, so this switches it to weights_only=True and adds
focused regression coverage for both model loaders.
Constraint: Must preserve compatibility with tensor-only legacy LoRA checkpoints
Rejected: Remove .ckpt/.pth support entirely | too disruptive for existing users
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep LoRA artifact handling aligned with the existing safetensors-first, weights-only loading pattern
Tested: python3 -m pytest -q tests/test_lora_checkpoint_loading.py tests/test_model_utils.py -q
Not-tested: Full end-to-end LoRA hot-load with heavyweight model assets
VoxCPM checkpoints default to bfloat16. Following commit e4e0496 which
added MPS device routing, running with `device=mps` selects bf16 on
Apple Silicon. On Metal, bf16 introduces enough numerical drift in the
diffusion AR loop that the synthesized audio is glitched and trips the
model's badcase detector, which retries until the per-call retry budget
is exhausted. Effectively MPS support is unusable in the default config.
This patch adds a single helper, `pick_runtime_dtype(device, dtype)`,
that promotes any low-precision dtype to float32 when the resolved
device is `mps`. CUDA and CPU paths are untouched. An opt-out env var
`VOXCPM_MPS_DTYPE` lets users force a specific dtype on MPS once future
PyTorch / macOS releases improve bf16 stability.
Both VoxCPMModel and VoxCPM2Model adopt the helper in their __init__,
replacing what would otherwise be duplicated inline checks.
Verified locally on Apple M5 Max, PyTorch 2.11, macOS 15:
- VoxCPM2 (2B): clean output, RTF ~0.78 steady state
- VoxCPM 0.5B: clean output, RTF ~0.92
- No badcase retries fired in any test
- VOXCPM_MPS_DTYPE=bfloat16 round-trips and reproduces the original
glitched output, confirming the override path.
Move generator close handling into a shared utility and wire the core generation pipeline through it so partially-consumed prompt cache generators are cleaned up consistently across both model variants and the public VoxCPM wrapper.
Made-with: Cursor
Factor repeated next-and-close patterns into a shared helper in both VoxCPM model variants so non-streaming inference cleans up generators consistently while keeping the issue reference close to the workaround.
Made-with: Cursor
Use context managers when reading config.json in VoxCPMModel.from_local()
and VoxCPM2Model.from_local() to prevent file descriptor leaks. Also add
explicit encoding="utf-8" to avoid locale-dependent decode errors.
Closes#235
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Support optional ref_audio samples in finetuning and make runtime device selection explicit while keeping auto fallback behavior consistent. Also ignore the local app override file to avoid accidental commits.
Made-with: Cursor
ZGY