Add Apple Silicon (MPS) inference support [claude]

[mparrett]

Summary

• Plumb use_fp16 through Decoder → Generator → SourceModuleHnNSF to support fp16 inference on MPS (fixes dtype mismatch when decoder runs in half precision)
• Add Demo/inference_mps.py — clean MPS inference script with -p flag for one-shot benchmarking
• Fix ref_texts → texts variable name bug in LibriTTS demo notebook
• Add pyproject.toml + lockfile for reproducible uv sync setup

MPS notes

TextEncoder must stay on CPU because MPS doesn't support pack_padded_sequence. All other modules run on MPS. After text encoding on CPU, tensors are transferred to the GPU device.

Scope is inference only — the training path has hardcoded .to('cuda') in the Decoder forward method which is a separate fix.

Usage

uv sync
USE_MPS=1 uv run python Demo/inference_mps.py -p "Hello world"

# With fp16 decoder (faster on some hardware)
USE_MPS=1 USE_FP16=1 uv run python Demo/inference_mps.py -p "Hello world"

Benchmarks (Apple Macbook Air M2, ~60 word passage → 17.7s audio)

Config	Inference	RTF	vs CPU
CPU	6.1s	0.36	—
MPS	4.4s	0.25	1.4x faster
MPS + FP16	2.9s	0.16	2.1x faster

RTF = real-time factor (lower is better). MPS+FP16 sustains ~5x real-time synthesis.

If the supply of fruit is greater than the family needs, it may be made a source of income by sending the fresh fruit to the market if there is one
near enough, or by preserving, canning, and making jelly for sale. To make such an enterprise a success the fruit and work must be first class.

Test plan

• Synthesized audio on Apple Silicon M1 with USE_MPS=1
• Verify fp16 path with USE_MPS=1 USE_FP16=1
• Verify CUDA path still works (no regressions — default use_fp16=False leaves all paths unchanged)
• Verify istftnet decoder path is unaffected (no changes to that module)

🤖 Generated with Claude Code