dgrauet/ernie-image-sft-mlx-q8

Int8 quantization (group_size 64, attention and FFN Linear weights only, leaving the small projections and the VAE in bf16) of dgrauet/ernie-image-sft-mlx, the MLX conversion of baidu/ERNIE-Image.

Quantized with mlx-forge (mlx-forge convert ernie-image --variant sft --quantize --bits 8).

Usage

These weights can be used with ernie-image-mlx:

pip install ernie-image-mlx
ernie-image-mlx generate \
    -p "一只黑白相间的中华田园犬" \
    --repo-id dgrauet/ernie-image-sft-mlx-q8 \
    -o dog.png

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

• mlx-forge (conversion): https://github.com/dgrauet/mlx-forge
• mlx-arsenal (reusable MLX ops): https://github.com/dgrauet/mlx-arsenal
• mlx-porting (Claude Code skill): https://github.com/dgrauet/claude-skill-mlx-porting
• bf16 variant: https://huggingface.co/dgrauet/ernie-image-sft-mlx
• q4 variant: https://huggingface.co/dgrauet/ernie-image-sft-mlx-q4

Files

• model_index.json (547.00 B)
• pe_tokenizer_tokenizer_config.json (20.63 KB)
• quantize_config.json (107.00 B)
• scheduler_scheduler_config.json (482.00 B)
• special_tokens_map.json (414.00 B)
• split_model.json (1.07 KB)
• text_encoder.safetensors (3.39 GB)
• text_encoder_config.json (1.64 KB)
• tokenizer.json (8.83 MB)
• tokenizer_tokenizer_config.json (379.00 B)
• transformer.safetensors (8.11 GB)
• transformer_config.json (383.00 B)
• vae.safetensors (160.33 MB)
• vae_config.json (831.00 B)