camilbradea/sam-audio-worker repository overview

⁠SAM-Audio GPU Worker

RunPod-ready container for Meta's SAM-Audio⁠ audio separation model.

⁠Contents

• Base: nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04
• Python 3.11 with PyTorch, torchaudio, xformers
• SAM-Audio + dependencies (CLAP, ImageBind, DAC-VAE, perception-models)
• FastAPI handler on port 8000
• SSH server (optional, for debugging)

⁠Optimizations

Uses a "lite" mode (inspired by AudioGhostAI⁠) that strips unused video components to reduce VRAM:

• Removed vision_encoder (~2GB saved)
• Removed visual_ranker (~2GB saved)
• Removed text_ranker (~2GB saved)

Audio-only processing - video input not supported.

⁠Authentication

All endpoints require X-API-Key header.

⁠API Endpoints

⁠POST /process

Start an audio separation job.

Request:

{
  "job_id": "uuid-string",
  "user_id": "user-identifier",
  "prompt": "drums",
  "original_filename": "song.wav",
  "use_float32": true,
  "timeout_seconds": 43200
}

Response:

{
  "status": "processing",
  "job_id": "uuid-string"
}

⁠GET /status/{job_id}

Poll job progress.

Response:

{
  "job_id": "uuid-string",
  "status": "processing|completed|failed",
  "output_key": "user-id/outputs/song_drums_residual.wav",
  "error": null,
  "chunks_processed": 3,
  "total_chunks": 10,
  "last_progress_ts": 1708012345.123
}

⁠GET /health

Response:

{
  "status": "ok",
  "model_loaded": true,
  "active_jobs": 1
}

⁠GET /startup-status

Response:

{
  "state": "starting|loading_model|ready|error",
  "error": null
}

⁠Requirements

• NVIDIA GPU with CUDA 12.4+ support
• ~32GB VRAM recommended

⁠Environment Variables

⁠Usage

docker run --gpus all -p 8000:8000 \
  -e R2_ACCOUNT_ID=xxx \
  -e R2_ACCESS_KEY=xxx \
  -e R2_SECRET_KEY=xxx \
  -e R2_BUCKET=xxx \
  -e API_KEY=xxx \
  -e HF_TOKEN=xxx \
  bradeac/sam-audio-worker:latest