fix(mps): force float32 on Apple Silicon to avoid bf16 quality loss

[Oumnya]

Summary

Following commit e4e0496 which added MPS device routing, running with device=mps selects the checkpoint default bfloat16 on Apple Silicon. On Metal, bf16 introduces enough numerical drift in the diffusion AR loop that the synthesized audio is noticeably glitched and consistently trips the model's badcase detector, which then retries until the per-call retry budget is exhausted. In effect, MPS support is unusable in the default config — users either get garbled audio or a hung process.

This PR adds a single helper, pick_runtime_dtype(device, configured_dtype), in model/utils.py 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 advanced users force a specific dtype on MPS for future testing as PyTorch / macOS Metal bf16 support improves.

Both VoxCPMModel and VoxCPM2Model __init__ adopt the helper, replacing what would otherwise be duplicated inline checks.

Reproducer (before this patch)

from voxcpm import VoxCPM
m = VoxCPM.from_pretrained("openbmb/VoxCPM2", load_denoiser=False, optimize=False, device="mps")
wav = m.generate(text="Hello from Apple Silicon.", cfg_value=2.0, inference_timesteps=10)

Observed on Apple M5 Max, macOS 15, PyTorch 2.11:

• Audio output is glitched / unintelligible
• Badcase detected, audio_text_ratio=... retries fire repeatedly and exhaust retry_badcase_max_times
• RTF inflated by ~2x due to retries even when output finally completes

Behavior table

device	configured dtype	resolved dtype	notes
cpu	bfloat16	bfloat16	unchanged
cuda	bfloat16	bfloat16	unchanged
mps	bfloat16	float32	new
mps	float16	float32	new
mps	float32	float32	unchanged
mps (with VOXCPM_MPS_DTYPE=bfloat16)	any	bfloat16	opt-out
mps (with VOXCPM_MPS_DTYPE=banana)	any	raises ValueError	invalid

Verified locally

Apple M5 Max, 128 GB unified memory, PyTorch 2.11.0, macOS 15.

Model	Device	Outcome	Steady-state RTF
VoxCPM 0.5B	cpu	clean	1.45
VoxCPM 0.5B	mps (bf16, before)	glitched + badcase loop	1.73
VoxCPM 0.5B	mps (fp32, after)	clean	0.92
VoxCPM2 2B	mps (fp32, after)	clean	0.78

Voice cloning, Voice Design, and Ultimate Cloning paths all verified end-to-end. Multilingual generation (12 languages) verified. No badcase retries fired in any post-patch run.

The VOXCPM_MPS_DTYPE=bfloat16 override round-trips: it reproduces the original glitched output, confirming the env path works and that the dtype is the actual cause.

Notes for review

• The helper is intentionally minimal (single function, dtype constants, env override) so it is easy to extend if more devices need similar handling later.
• I kept the call sites in VoxCPMModel.__init__ and VoxCPM2Model.__init__ rather than burying the promotion inside setup_cache so the user-facing log clearly shows the effective dtype on startup.
• No README changes in this PR — happy to add a short Apple Silicon section in a follow-up if maintainers want it.

[a710128]

I noticed a small inconsistency in the new MPS override path.

pick_runtime_dtype() accepts VOXCPM_MPS_DTYPE=half, but get_dtype() does not currently recognize "half" and will later raise ValueError: Unsupported dtype: half.

So the override validation is a bit broader than the actual dtype parser. It looks like this could be fixed either by:

• adding "half" support in get_dtype(), or
• removing "half" from the accepted override values.

Not blocking from a design point of view, but I think this should be aligned before merge to avoid a confusing runtime failure for users trying the override.

[a710128]

Looks good to me — the previous half inconsistency is fixed, and the added guard makes the dtype override behavior much clearer.