AUser0's picture
Update README.md
d86c870 verified
metadata
license: mit
base_model:
  - meta-llama/Llama-3.2-3B
tags:
  - custom-kernel
  - llama
  - quantization

Llama-3.2-3B (SpectralQ 6-bit) This model contains experimental 6-bit frequency-domain weights quantized using the Transformer Spectral Scaling Law. It utilizes 256-width DCT macroblocks and a 84-bit chunk-major packed layout.

⚠️ IMPORTANT: HOW TO USE ⚠️ This model cannot be loaded with standard Transformers or vLLM. It requires the custom SpectralQ inference engine, which includes the C++ SRAM bit-unpacking and IDCT projection kernels.

To run this model, please visit the GitHub repository: [👉 SpectralQ 👈]