fix: make SoulX-Podcast run on macOS (MPS/CPU)

[gu20241212-crypto]

Summary

• make inference device selection adaptive (cuda / mps / cpu) instead of hardcoded CUDA
• remove CUDA-only tensor/model moves in SoulXPodcast core pipeline
• make HF engine dtype/device initialization compatible with Apple Silicon
• auto-fallback to hf engine when vllm/CUDA is unavailable
• guard CUDA seed call in random-seed utility
• add requirements.macos.txt for macOS-compatible dependency install

Why

Current code path fails on macOS (Apple Silicon) due to CUDA hardcoding (.cuda(), CUDA-only assumptions in engine setup), causing runtime errors such as Torch not compiled with CUDA enabled.

Validation

• local syntax check passed via python -m py_compile
• end-to-end TTS generation succeeded on macOS using:
  • cli/tts.py with model pretrained_models/SoulX-Podcast-1.7B
  • output wav files generated successfully (24kHz PCM)

[gu20241212-crypto]

Validation notes (macOS Apple Silicon):

• Environment: macOS + Apple Silicon (MPS available, CUDA unavailable).
• Goal: verify this patch fixes CUDA hardcoding and enables successful inference on macOS.

Commands used:

cd /Users/GUGU/SoulX_Podcast/SoulX-Podcast
export PYTHONPATH=$PWD
/Users/GUGU/SoulX_Podcast/.venv/bin/python cli/tts.py \
  --prompt_text "喜欢攀岩、徒步、滑雪的语言爱好者。" \
  --dialect_prompt "" \
  --prompt_audio example/audios/female_mandarin.wav \
  --text "你好，这是一次短句测试。" \
  --model_path pretrained_models/SoulX-Podcast-1.7B \
  --output_path outputs/mandarin_tts_macos_fix_short.wav \
  --seed 7

Observed result:

• Inference completed successfully.
• Output wav written: outputs/mandarin_tts_macos_fix_short.wav
• Audio info: 24kHz, mono, PCM_16, ~2.80s.

Also verified a longer sample generation completed successfully (~11.36s output).

Compatibility expectation:

• CUDA path remains available when CUDA exists.
• On non-CUDA devices (e.g. macOS/MPS), runtime now uses adaptive device selection and avoids CUDA-only code paths.