Introducing EntroPack: a PyTorch tensor compression library that lets you tune weight storage to your GPU memory budget.
🤖 https://modelscope.ai/models/DiffSynth-Studio/EntroPackPreQuants
🎛️ Flexible compression: choose fractional bits per parameter instead of fixed 4-bit or 8-bit formats. No activation calibration or fine-tuning required.
🏆 Better use of 16 GB VRAM: in a Qwen-Image-2.1 test at 1024 x 1024, targeting 4.8 bpp peaked at 15.86 GiB while cutting weight reconstruction error from NF4's 9.40% to 4.26%.
⚡ GPU-native decoding: E8 lattice quantization plus entropy coding, with fused decoding and reconstruction in a single kernel. The reported BF16 single-tensor test exceeds 340 GB/s.
📦 Ready to try: pre-quantized weights for Z-Image-Turbo, Qwen-Image-2.1, and MiniMax-H3 at 4 to 8 bpp, for DiffSynth-Studio and ComfyUI. The extreme Z-Image build reaches 2.3 bpp, shrinking its DiT weights to 1.6 GB, around one-seventh of BF16.
📜 Library code under Apache 2.0.
💻 Code: https://github.com/modelscope/entropack
📄 Technical report: https://modelscope.ai/papers/2609.34185