CogVideoX-Fun-V1.5-5b-InP-mlx-q8

dgrauet/CogVideoX-Fun-V1.5-5b-InP-mlx-q8

Int8 quantization (group_size 64, transformer block Linear weights only, leaving the embeddings, norms and the output projection in bf16) of dgrauet/CogVideoX-Fun-V1.5-5b-InP-mlx, the MLX conversion of alibaba-pai/CogVideoX-Fun-V1.5-5b-InP.

Quantized with mlx-forge (mlx-forge convert cogvideox-fun-v1.5-5b-inp --quantize --bits 8).

Usage

These weights can be used with VideoX-Fun-mlx:

pip install mlx sentencepiece pillow numpy huggingface_hub
pip install git+https://github.com/dgrauet/mlx-arsenal.git
git clone https://github.com/dgrauet/VideoX-Fun-mlx.git && cd VideoX-Fun-mlx

huggingface-cli download dgrauet/CogVideoX-Fun-V1.5-5b-InP-mlx-q8 --local-dir models/cogvideox-fun

python scripts/quick_infer.py \
    --model-path models/cogvideox-fun \
    --prompt "a beautiful sunset over the ocean" \
    --output sunset.gif

Keep quantize_config.json next to the weights (the loader also infers bits/group_size from the weight shapes if it is missing).

Related Projects

• VideoX-Fun-mlx (inference code): https://github.com/dgrauet/VideoX-Fun-mlx
• mlx-forge (conversion tool): https://github.com/dgrauet/mlx-forge
• mlx-arsenal (MLX utilities): https://github.com/dgrauet/mlx-arsenal
• bf16 variant: https://huggingface.co/dgrauet/CogVideoX-Fun-V1.5-5b-InP-mlx
• q4 variant: https://huggingface.co/dgrauet/CogVideoX-Fun-V1.5-5b-InP-mlx-q4

Files

• config.json (2.78 KB)
• configuration.json (56.00 B)
• model_index.json (411.00 B)
• quantize_config.json (129.00 B)
• scheduler_scheduler_config.json (482.00 B)
• spiece.model (773.10 KB)
• split_model.json (1.56 KB)
• text_encoder.safetensors (8.87 GB)
• text_encoder_config.json (782.00 B)
• tokenizer_added_tokens.json (2.53 KB)
• tokenizer_special_tokens_map.json (2.48 KB)
• tokenizer_tokenizer_config.json (20.13 KB)
• transformer.safetensors (6.22 GB)
• transformer_config.json (887.00 B)
• vae.safetensors (411.25 MB)
• vae_config.json (839.00 B)