# Vocence chute config for MOSS-TTS ~18B + example_repo/miner.py (build-time).
# Weights live in the HF repo snapshot; this image supplies PyTorch and deps only.
#
# ``miner.py`` does not import ``torch``/``transformers`` at module load (Vocence runs an
# import audit during ``exec_module``). Weights load on first ``warmup()``/``generate_wav()``,
# after stripping ``VocenceSandboxImporter``. Device defaults use ``/dev/nvidia0`` unless
# ``runtime.device`` is set in ``vocence_config.yaml`` or ``FORCE_CPU=1``.

Image:
  from_base: parachutes/base-python:3.12.9
  set_user: root
  run_command:
    # soundfile (used by miner’s torchaudio.load patch) needs libsndfile at runtime.
    - apt-get update && apt-get install -y --no-install-recommends libsndfile1 && rm -rf /var/lib/apt/lists/*
    # GPU wheels; if your Chutes stack uses another CUDA major, switch cu124 → cu126/cu128 on pytorch.org.
    - pip install --no-cache-dir torch torchaudio --index-url https://download.pytorch.org/whl/cu124
    - pip install --no-cache-dir transformers accelerate huggingface_hub safetensors pyyaml soundfile
    - pip install --no-cache-dir numpy einops scipy tqdm protobuf sentencepiece filelock packaging regex requests

NodeSelector:
  # 18B bf16 weights are ~36GB; reserve headroom for activations / tokenizer / decode.
  gpu_count: 2
  include: ['pro_6000']
  min_vram_gb_per_gpu: 120
  exclude: []

Chute:
  tagline: Vocence PromptTTS — MOSS-TTS 18B (miner.py)
  readme: MOSS-TTS 18B + Vocence miner; single GPU, high-VRAM nodes (see min_vram_gb_per_gpu).
  shutdown_after_seconds: 86400
  concurrency: 1
  max_instances: 1
  scaling_threshold: 0.5
  tee: true