Instructions to use FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree-MLX-INT4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree-MLX-INT4 with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree-MLX-INT4 FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree-MLX-INT4
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
| { | |
| "schema_version": "fastvideo-fasth3-mlx-conversion-v1", | |
| "published_repo_id": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree-MLX-INT4", | |
| "source": { | |
| "repo_id": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree", | |
| "revision": "f624f08c6c279ab43534c003e556fc5b295b6558", | |
| "checkpoint_step": 1000, | |
| "attention": "dense", | |
| "checkpoint_content_aggregate_sha256": "3f56129acb9749f2ae4e52bab176c09406bf027d68c6f020cf5ff99e651b1609", | |
| "transformer_shards_verified": 13 | |
| }, | |
| "converter": { | |
| "fastvideo_commit": "cf6a00b9be4675602126d6aeab902ad9a74810ea", | |
| "script": "scripts/checkpoint_conversion/convert_minimax_h3_mlx.py", | |
| "command": "python scripts/checkpoint_conversion/convert_minimax_h3_mlx.py --model-root <snapshot>/transformer --out <output> --formats 'int8 int6 int4'", | |
| "mlx_version": "0.32.2", | |
| "python_version": "3.12.13", | |
| "hardware": "Apple M4 Max", | |
| "unified_memory_bytes": 38654705664, | |
| "combined_wall_seconds": 164.62, | |
| "process_max_rss_bytes": 11317051392, | |
| "peak_memory_footprint_bytes": 27108979504 | |
| }, | |
| "format": { | |
| "conversion_seconds": 46.1, | |
| "peak_mlx_gib": 14.8, | |
| "weights_file": "mlx_h3_dit.safetensors", | |
| "weights_bytes": 11532219083, | |
| "weights_sha256": "dcee86c17197c621a0e3d3ba9541cb6ef5764cd0c258beb7b044fb12b217e948", | |
| "manifest_file": "mlx_h3_dit.json", | |
| "manifest_sha256": "25c89f37f53292c4abd8bbb70098eaad54783a1b1fc5f0f55ed779a8c85efb56", | |
| "quantization": { | |
| "mode": "affine", | |
| "bits": 4, | |
| "group_size": 64 | |
| }, | |
| "tensor_count": 1464, | |
| "vsa_capable": false | |
| }, | |
| "validation": { | |
| "source_shard_size_and_sha256": "passed, 13 of 13", | |
| "safetensors_header_open": "passed", | |
| "manifest_quantization": "passed", | |
| "adaln_cache_timesteps": 8, | |
| "generation": "passed: 832x480, 124 frames, full H3 VAE", | |
| "runtime_checkpoint_load": "passed for int8, int6, and int4", | |
| "runtime_structure": "50 transformer blocks, 2 refiner blocks, 50 AdaLN block tables" | |
| } | |
| } | |