vanch007/FastVideo-FastH3-8-Step-V2-MLX-INT6

FastVideo FastH3-8-Step-V2 (MLX INT6 Quantized with VSA)

This repository contains the MLX-native INT6 quantized checkpoint of FastVideo/FastVideo-FastH3-8-Step-V2, fully optimized for Apple Silicon Macs (M2/M3/M4 Max/Ultra).

Key Highlights

• Format: MLX affine quantized INT6 (group_size=64, bits=6) with VSA (Variable Sparse Attention) support.
• Size: ~17 GB safetensors (reduced from original 65+ GB BF16).
• Quality: Verified 100% visual parity with original BF16 weights on Apple Silicon.
• Performance: End-to-end inference takes ~29 minutes on Apple M3 Max (128GB) with high-quality tiled VAE decoding (8 denoising steps).

---

How to Use

Method 1: Python API via FastVideo

Clone and install FastVideo:

Run inference with Python:

---

Method 2: Command Line Interface (CLI)

Run directly using FastVideo's pipeline runner:

---

Method 3: Direct Hub Download & Local Deployment

Download and place the checkpoint files into your FastVideo directory:

/Users/vanch/FastVideo/models/FastH3-MLX-vsa/int6

Ensure you have downloaded the base pipeline components (text_encoder, vae, processor, etc.) from the official FastVideo/FastVideo-FastH3-8-Step-V2 repo.

---

Hardware Recommendations

• Recommended: Apple M3/M4 Max with 64GB or 128GB Unified Memory.
• Power: Keep power adapter connected and enable "High Power Mode" in macOS Battery Settings for optimal sustained GPU throughput.