gashapon/Qwen3.8-27B-ROCmFPX

<p align="center"> <img src="./assets/logo.png" width="200" alt="Logo"> </p>

These quantized models can't be run from mainstream llama.cpp. In order to use them, you should use a fork that supports the ROCmFPX format (e.g., ROCmFPX, LaurentZuijdwijk/llama.cpp)

This repository contains the majority of the ROCmFPX quantizations for Qwen3.8-27B (quantized from BF16). It also includes an imatrix file generated using Qwen3.8-Q8_0 (unsloth/Qwen3.8-27B-GGUF) and the following calibration data: bartowski1182/calibration_datav5-txt.

Quantization methodology used

1. 1.Generating the matrix:

$ llama-imatrix \
    -m Qwen3.8-27B-Q8_0.gguf \
    -f calib.txt \
    -o qwen3-27b.imatrix \
    -c 512

1. 1.Generating the quantizations:

$ llama-quantize \
    --imatrix qwen3-27b.imatrix \
    Qwen3.8-27B-BF16-00001-of-00002.gguf \
    {output_name}.gguf \
    {quant_type} # list: https://github.com/charlie12345/ROCmFPX#which-format-should-i-pick

For further information, check the ROCmFPX repository!