| Original model: https://huggingface.co/brucethemoose/Yi-34B-200K-RPMerge | |
| Steps: | |
| 1. Convert to GGUF using llama.cpp (clone from source, install requirements, then run this) | |
| > `python convert.py /mnt/d/LLM_Models/Yi-34B-200K-RPMerge/ --vocab-type hfft --outtype f32 --outfile Yi-34B-200K-RPMerge.gguf` | |
| 2. Create imatrix (offload as much as you can to the GPU) | |
| > `./imatrix -m /mnt/d/LLM_Models/Yi-34B-200K-RPMerge.gguf -f /mnt/d/LLM_Models/8k_random_data.txt -o /mnt/d/LLM_Models/Yi-34B-200K-RPMerge.imatrix.dat -ngl 20` | |
| 3. Quantize using imatrix | |
| > `./quantize --imatrix /mnt/d/LLM_Models/Yi-34B-200K-RPMerge.imatrix.dat /mnt/d/LLM_Models/Yi-34B-200K-RPMerge.gguf /mnt/d/LLM_Models/Yi-34B-200K-RPMerge.IQ2_XXS.gguf IQ2_XXS | |
| I have also uploaded [8k_random_data.txt from this github discussion](https://github.com/ggerganov/llama.cpp/discussions/5006) | |
| And the importance matrix I made (`Yi-34B-200K-RPMerge.imatrix.dat`) |