sam-audio-infer 0.1.1

Optimized inference package for Meta's SAM-Audio model with VRAM-efficient lite mode

sam-audio-infer

PyPI: https://pypi.org/project/sam-audio-infer/

Inference-only package for SAM-Audio (Segment Anything for Audio) by Meta AI.

This is a lightweight, dependency-minimal repackaging focused solely on inference with VRAM-efficient lite mode. For training and the full research codebase, please visit the original SAM-Audio repository.

---

Features

• Inference-Only: Optimized for inference with torch.inference_mode() (no grad overhead)
• Lite Mode: Reduce VRAM usage by 62-78% by removing unused components
• Mixed Precision: Support for bfloat16/float16 inference
• 48kHz Audio: Native high-quality audio processing at 48kHz sample rate
• Auto-Chunking: Process long audio files without OOM errors
• Model Caching: Configurable cache directory with environment variable support
• Warmup Support: Pre-compile CUDA kernels for faster first inference
• Simple API: Easy-to-use Python API and CLI

Installation

# Using uv (recommended)
uv add sam-audio-infer
uv pip install git+https://github.com/facebookresearch/sam-audio.git

# Or using pip
pip install sam-audio-infer
pip install git+https://github.com/facebookresearch/sam-audio.git

For development:

git clone https://github.com/openmirlab/sam-audio-infer.git
cd sam-audio-infer
uv sync

Prerequisites

HuggingFace Access Required: SAM-Audio models are gated.

1. Request access to the model checkpoints:
   • facebook/sam-audio-base
   • facebook/sam-audio-large
2. Once accepted, authenticate with HuggingFace:

   # Generate token at https://huggingface.co/settings/tokens
   huggingface-cli login
   # Or set environment variable
   export HF_TOKEN=hf_your_token_here
   

Quick Start

Python API

from sam_audio_infer import SamAudioInfer

# Load model (recommended settings, ~3 GB VRAM)
model = SamAudioInfer.from_pretrained(
    "base",                      # Model size: "small", "base", or "large"
    dtype="bfloat16",            # Mixed precision (~50% VRAM savings)
    enable_text_ranker=False,    # +3 GB VRAM if enabled
    enable_span_predictor=False, # +3 GB VRAM if enabled
)

# Separate audio
result = model.separate("song.wav", description="vocals")
result.save("vocals.wav", "accompaniment.wav")

Command Line

# Basic separation
sam-audio-infer separate song.wav -d "vocals" -o vocals.wav

# With residual output
sam-audio-infer separate song.wav -d "drums" -o drums.wav --residual other.wav

# Download model with warmup
sam-audio-infer download --model base --warmup

VRAM Usage

Model	Full Mode	Lite Mode	Reduction
Base	12.73 GB	2.84 GB	78%
Large	16.18 GB	6.15 GB	62%

Tested on RTX 4090 with bfloat16 precision

Documentation

• CLI Reference - Command line interface
• Python API - Python API reference
• Configuration - Models, precision, lite mode settings
• Architecture - How it works and optimization techniques
• Benchmarks - VRAM and performance benchmarks
• Troubleshooting - Common issues and solutions

Requirements

• Python >= 3.11
• PyTorch >= 2.0.0
• torchaudio >= 2.0.0
• CUDA-capable GPU with at least 4GB VRAM (lite + bfloat16)

Acknowledgments

This package stands on the shoulders of two important projects.

Original Research by Meta AI / Facebook Research

SAM-Audio (Segment Anything for Audio) is developed by Meta AI Research.

• Repository: github.com/facebookresearch/sam-audio
• Paper: Segment Anything for Audio
• HuggingFace: facebook/sam-audio-base

Lite Mode Optimization

The Lite Mode VRAM optimization technique used in this package is inspired by AudioGhost AI.

License

MIT License

Note: The underlying SAM-Audio model has its own license terms. Please refer to the official SAM-Audio repository for model usage terms.

Citation

If you use SAM Audio in your research, please cite the original paper:

@article{shi2025samaudio,
    title={SAM Audio: Segment Anything in Audio},
    author={Bowen Shi and Andros Tjandra and John Hoffman and Helin Wang and Yi-Chiao Wu and Luya Gao and Julius Richter and Matt Le and Apoorv Vyas and Sanyuan Chen and Christoph Feichtenhofer and Piotr Doll{\'a}r and Wei-Ning Hsu and Ann Lee},
    year={2025},
    url={https://arxiv.org/abs/2512.18099}
}