quant-clone 0.1.1

Generate a llama-quantize command to copy the quantization parameters of any GGUF

quant_clone

This is a simple little script to help you generate a llama-quantize (from llama.cpp) command which will allow you to quantize your own GGUF the same way your target GGUF has been quantized.

Installation

pip install quant_clone

if the published gguf package doesn't support your model yet, install the current one with:

pip install --force-reinstall --upgrade "git+https://github.com/ggml-org/llama.cpp.git#egg=gguf&subdirectory=gguf-py"

Usage

quant_clone input.gguf output.txt

input.gguf is the GGUF file whose quantization parameters you would like to copy

output.txt parameter is optional, if it's omitted the output will be saved to cmd.txt

Example

if I take one of unsloth's dynamic 2.0 quants and run:

quant_clone gemma-3-1b-it-UD-IQ1_S.gguf

I get this output:

llama-quantize --imatrix <imatrix_unsloth.dat> --tensor-type token_embd.weight=Q5_1 --tensor-type blk.0.attn_k.weight=IQ4_NL --tensor-type blk.0.attn_output.weight=IQ2_XXS --tensor-type blk.0.attn_q.weight=IQ4_NL --tensor-type blk.0.attn_v.weight=Q5_0 --tensor-type blk.0.ffn_down.weight=IQ3_S --tensor-type blk.0.ffn_gate.weight=IQ4_NL --tensor-type blk.0.ffn_up.weight=IQ4_NL --tensor-type blk.1.attn_k.weight=IQ4_NL --tensor-type blk.1.attn_output.weight=IQ2_XXS --tensor-type blk.1.attn_q.weight=IQ4_NL --tensor-type blk.1.attn_v.weight=Q5_0 --tensor-type blk.1.ffn_down.weight=Q2_K --tensor-type blk.1.ffn_gate.weight=IQ4_NL --tensor-type blk.1.ffn_up.weight=IQ4_NL --tensor-type blk.2.attn_k.weight=IQ4_NL --tensor-type blk.2.attn_output.weight=IQ2_XXS --tensor-type blk.2.attn_q.weight=IQ4_NL --tensor-type blk.2.attn_v.weight=Q5_0 --tensor-type blk.2.ffn_down.weight=IQ3_S --tensor-type blk.2.ffn_gate.weight=IQ4_NL --tensor-type blk.2.ffn_up.weight=IQ4_NL --tensor-type blk.3.attn_k.weight=IQ4_NL --tensor-type blk.3.attn_output.weight=IQ2_XXS --tensor-type blk.3.attn_q.weight=IQ4_NL --tensor-type blk.3.attn_v.weight=Q5_0 --tensor-type blk.3.ffn_down.weight=IQ3_S --tensor-type blk.3.ffn_gate.weight=IQ4_NL --tensor-type blk.3.ffn_up.weight=IQ4_NL --tensor-type blk.4.attn_k.weight=IQ4_NL --tensor-type blk.4.attn_output.weight=IQ2_XXS --tensor-type blk.4.attn_q.weight=IQ4_NL --tensor-type blk.4.attn_v.weight=Q5_0 --tensor-type blk.4.ffn_down.weight=IQ3_S --tensor-type blk.4.ffn_gate.weight=IQ4_NL --tensor-type blk.4.ffn_up.weight=IQ4_NL --tensor-type blk.5.attn_k.weight=IQ4_NL --tensor-type blk.5.attn_output.weight=IQ2_XXS --tensor-type blk.5.attn_q.weight=IQ4_NL --tensor-type blk.5.attn_v.weight=Q5_0 --tensor-type blk.5.ffn_down.weight=IQ1_S --tensor-type blk.5.ffn_gate.weight=IQ4_NL --tensor-type blk.5.ffn_up.weight=IQ4_NL --tensor-type blk.6.attn_k.weight=IQ4_NL --tensor-type blk.6.attn_output.weight=IQ2_XXS --tensor-type blk.6.attn_q.weight=IQ4_NL --tensor-type blk.6.attn_v.weight=Q5_0 --tensor-type blk.6.ffn_down.weight=IQ1_S --tensor-type blk.6.ffn_gate.weight=IQ4_NL --tensor-type blk.6.ffn_up.weight=IQ4_NL --tensor-type blk.7.attn_k.weight=IQ4_NL --tensor-type blk.7.attn_output.weight=IQ2_XXS --tensor-type blk.7.attn_q.weight=IQ4_NL --tensor-type blk.7.attn_v.weight=Q5_0 --tensor-type blk.7.ffn_down.weight=IQ1_S --tensor-type blk.7.ffn_gate.weight=IQ4_NL --tensor-type blk.7.ffn_up.weight=IQ4_NL --tensor-type blk.8.attn_k.weight=IQ4_NL --tensor-type blk.8.attn_output.weight=IQ2_XXS --tensor-type blk.8.attn_q.weight=IQ4_NL --tensor-type blk.8.attn_v.weight=Q5_0 --tensor-type blk.8.ffn_down.weight=IQ1_S --tensor-type blk.8.ffn_gate.weight=IQ4_NL --tensor-type blk.8.ffn_up.weight=IQ4_NL --tensor-type blk.9.attn_k.weight=IQ4_NL --tensor-type blk.9.attn_output.weight=IQ2_XXS --tensor-type blk.9.attn_q.weight=IQ4_NL --tensor-type blk.9.attn_v.weight=Q5_0 --tensor-type blk.9.ffn_down.weight=IQ1_S --tensor-type blk.9.ffn_gate.weight=IQ4_NL --tensor-type blk.9.ffn_up.weight=IQ4_NL --tensor-type blk.10.attn_k.weight=IQ4_NL --tensor-type blk.10.attn_output.weight=IQ2_XXS --tensor-type blk.10.attn_q.weight=IQ4_NL --tensor-type blk.10.attn_v.weight=Q5_0 --tensor-type blk.10.ffn_down.weight=IQ1_S --tensor-type blk.10.ffn_gate.weight=IQ4_NL --tensor-type blk.10.ffn_up.weight=IQ4_NL --tensor-type blk.11.attn_k.weight=IQ4_NL --tensor-type blk.11.attn_output.weight=IQ2_XXS --tensor-type blk.11.attn_q.weight=IQ4_NL --tensor-type blk.11.attn_v.weight=Q5_0 --tensor-type blk.11.ffn_down.weight=IQ2_S --tensor-type blk.11.ffn_gate.weight=IQ4_NL --tensor-type blk.11.ffn_up.weight=IQ4_NL --tensor-type blk.12.attn_k.weight=IQ4_NL --tensor-type blk.12.attn_output.weight=IQ2_XXS --tensor-type blk.12.attn_q.weight=IQ4_NL --tensor-type blk.12.attn_v.weight=Q5_0 --tensor-type blk.12.ffn_down.weight=IQ2_S --tensor-type blk.12.ffn_gate.weight=IQ4_NL --tensor-type blk.12.ffn_up.weight=IQ4_NL --tensor-type blk.13.attn_k.weight=IQ4_NL --tensor-type blk.13.attn_output.weight=IQ2_XXS --tensor-type blk.13.attn_q.weight=IQ4_NL --tensor-type blk.13.attn_v.weight=Q5_0 --tensor-type blk.13.ffn_down.weight=IQ2_S --tensor-type blk.13.ffn_gate.weight=IQ4_NL --tensor-type blk.13.ffn_up.weight=IQ4_NL --tensor-type blk.14.attn_k.weight=IQ4_NL --tensor-type blk.14.attn_output.weight=IQ2_XXS --tensor-type blk.14.attn_q.weight=IQ4_NL --tensor-type blk.14.attn_v.weight=Q5_0 --tensor-type blk.14.ffn_down.weight=IQ2_S --tensor-type blk.14.ffn_gate.weight=IQ4_NL --tensor-type blk.14.ffn_up.weight=IQ4_NL --tensor-type blk.15.attn_k.weight=IQ4_NL --tensor-type blk.15.attn_output.weight=IQ2_XXS --tensor-type blk.15.attn_q.weight=IQ4_NL --tensor-type blk.15.attn_v.weight=Q5_0 --tensor-type blk.15.ffn_down.weight=IQ2_S --tensor-type blk.15.ffn_gate.weight=IQ4_NL --tensor-type blk.15.ffn_up.weight=IQ4_NL --tensor-type blk.16.attn_k.weight=IQ4_NL --tensor-type blk.16.attn_output.weight=IQ2_XXS --tensor-type blk.16.attn_q.weight=IQ4_NL --tensor-type blk.16.attn_v.weight=Q5_0 --tensor-type blk.16.ffn_down.weight=IQ1_S --tensor-type blk.16.ffn_gate.weight=IQ4_NL --tensor-type blk.16.ffn_up.weight=IQ4_NL --tensor-type blk.17.attn_k.weight=IQ4_NL --tensor-type blk.17.attn_output.weight=IQ2_XXS --tensor-type blk.17.attn_q.weight=IQ4_NL --tensor-type blk.17.attn_v.weight=Q5_0 --tensor-type blk.17.ffn_down.weight=IQ1_S --tensor-type blk.17.ffn_gate.weight=IQ4_NL --tensor-type blk.17.ffn_up.weight=IQ4_NL --tensor-type blk.18.attn_k.weight=IQ4_NL --tensor-type blk.18.attn_output.weight=IQ2_XXS --tensor-type blk.18.attn_q.weight=IQ4_NL --tensor-type blk.18.attn_v.weight=Q5_0 --tensor-type blk.18.ffn_down.weight=IQ1_S --tensor-type blk.18.ffn_gate.weight=IQ4_NL --tensor-type blk.18.ffn_up.weight=IQ4_NL --tensor-type blk.19.attn_k.weight=IQ4_NL --tensor-type blk.19.attn_output.weight=IQ2_XXS --tensor-type blk.19.attn_q.weight=IQ4_NL --tensor-type blk.19.attn_v.weight=Q5_0 --tensor-type blk.19.ffn_down.weight=IQ1_S --tensor-type blk.19.ffn_gate.weight=IQ4_NL --tensor-type blk.19.ffn_up.weight=IQ4_NL --tensor-type blk.20.attn_k.weight=IQ4_NL --tensor-type blk.20.attn_output.weight=IQ2_XXS --tensor-type blk.20.attn_q.weight=IQ4_NL --tensor-type blk.20.attn_v.weight=Q5_0 --tensor-type blk.20.ffn_down.weight=IQ1_S --tensor-type blk.20.ffn_gate.weight=IQ4_NL --tensor-type blk.20.ffn_up.weight=IQ4_NL --tensor-type blk.21.attn_k.weight=IQ4_NL --tensor-type blk.21.attn_output.weight=IQ2_XXS --tensor-type blk.21.attn_q.weight=IQ4_NL --tensor-type blk.21.attn_v.weight=Q5_0 --tensor-type blk.21.ffn_down.weight=IQ1_S --tensor-type blk.21.ffn_gate.weight=IQ4_NL --tensor-type blk.21.ffn_up.weight=IQ4_NL --tensor-type blk.22.attn_k.weight=IQ4_NL --tensor-type blk.22.attn_output.weight=IQ2_XXS --tensor-type blk.22.attn_q.weight=IQ4_NL --tensor-type blk.22.attn_v.weight=Q5_0 --tensor-type blk.22.ffn_down.weight=IQ1_S --tensor-type blk.22.ffn_gate.weight=IQ4_NL --tensor-type blk.22.ffn_up.weight=IQ4_NL --tensor-type blk.23.attn_k.weight=IQ4_NL --tensor-type blk.23.attn_output.weight=IQ2_XXS --tensor-type blk.23.attn_q.weight=IQ4_NL --tensor-type blk.23.attn_v.weight=Q5_0 --tensor-type blk.23.ffn_down.weight=IQ1_S --tensor-type blk.23.ffn_gate.weight=IQ4_NL --tensor-type blk.23.ffn_up.weight=IQ4_NL --tensor-type blk.24.attn_k.weight=IQ4_NL --tensor-type blk.24.attn_output.weight=IQ2_XXS --tensor-type blk.24.attn_q.weight=IQ4_NL --tensor-type blk.24.attn_v.weight=Q5_0 --tensor-type blk.24.ffn_down.weight=IQ1_S --tensor-type blk.24.ffn_gate.weight=IQ4_NL --tensor-type blk.24.ffn_up.weight=IQ4_NL --tensor-type blk.25.attn_k.weight=IQ4_NL --tensor-type blk.25.attn_output.weight=IQ2_XXS --tensor-type blk.25.attn_q.weight=IQ4_NL --tensor-type blk.25.attn_v.weight=Q5_0 --tensor-type blk.25.ffn_down.weight=IQ3_S --tensor-type blk.25.ffn_gate.weight=IQ4_NL --tensor-type blk.25.ffn_up.weight=IQ4_NL <input.gguf> <output.gguf> Q8_0

That's the command to run to replicate the quantization. Make sure to edit imatrix path, input gguf path, and output gguf path.