base_model: google/gemma-2-9b-it
language:
- multilingual
datasets:
- TFMC/imatrix-dataset-for-japanese-llm
library_name: transformers
license: gemma
license_link: https://ai.google.dev/gemma/terms
pipeline_tag: text-generation
tags:
- nlp
- code
quantized_by: ymcki
widget:
- messages:
- role: user
content: Can you provide ways to eat combinations of bananas and dragonfruits?
Original model: https://huggingface.co/google/gemma-2-9b-it
Description
The purpose of this repository is to see whether Japanese specific imatrix can improve the performance of a non Japanese optimized model.
It also provides the Q4_0_8_8, Q4_0_4_8 and Q4_0_4_4 ggufs for edge devices that were otherwise not made by bartowski. These models should also be good for edge devices with 16GB RAM.
Prompt format
<start_of_turn>user
{prompt}<end_of_turn>
<start_of_turn>model
<end_of_turn>
<start_of_turn>model
Note that this model does not support a System prompt.
Download a file (not the whole branch) from below:
ELIZA-Tasks-100 is pretty standard benchmark for Japanese LLMs. The perfect score is 5.00. As a reference, bartowski's gemma-2-27b-it.Q6_K.gguf scores 4.04.
| Filename | Quant type | File Size | Split | ELIZA-Tasks-100 | Nvidia 3090 | Description |
|---|---|---|---|---|---|---|
| gemma-2-9b-it.f16.gguf | f16 | 18.49GB | false | 3.75 | 31.9t/s | Full F16 weights. |
| gemma-2-9b-it.Q8_0.gguf | Q8_0 | 9.83GB | false | 3.06 | 56.1t/s | Extremely high quality, recommended for edge devices with 16GB RAM. |
| gemma-2-2b-jpn-it-imatrix.Q4_0.gguf | Q4_0 | 1.63GB | false | 2.89 | 137t/s | Good quality, recommended for edge devices wth 8GB RAM. |
| gemma-2-2b-jpn-it-imatrix.Q4_0_8_8.gguf | Q4_0_8_8 | 1.63GB | false | TBD | TBD | Good quality, recommended for edge device <8GB RAM. |
| gemma-2-2b-jpn-it-imatrix.Q4_0_4_8.gguf | Q4_0_4_8 | 1.63GB | false | TBD | TBD | Good quality, recommended for edge device <8GB RAM. |
| gemma-2-2b-jpn-it-imatrix.Q4_0_4_4.gguf | Q4_0_4_4 | 1.63GB | false | TBD | TBD | Good quality, recommended for edge device <8GB RAM. |
| gemma-2-9b-it.Q4_0.gguf | Q4_0 | 5.44GB | false | 3.64 | 65.1t/s | Good quality, recommended for edge device with 8GB RAM |
| gemma-2-2b-jpn-it.Q4_0_8_8.gguf | Q4_0_8_8 | 1.63GB | false | TBD | TBD | Good quality, recommended for edge device <8GB RAM |
| gemma-2-2b-jpn-it.Q4_0_4_8.gguf | Q4_0_4_8 | 1.63GB | false | TBD | TBD | Good quality, recommended for edge device <8GB RAM |
| gemma-2-2b-jpn-it.Q4_0_4_4.gguf | Q4_0_4_4 | 1.63GB | false | TBD | TBD | Good quality, recommended for edge device <8GB RAM. |
How to check i8mm and sve support for ARM devices
ARM i8mm support is necessary to take advantage of Q4_0_4_8 gguf. All ARM architecture >= ARMv8.6-A supports i8mm.
ARM sve support is necessary to take advantage of Q4_0_8_8 gguf. sve is an optional feature that starts from ARMv8.2-A but majority of ARM chips doesn't implement it.
For ARM devices without both, it is recommended to use Q4_0_4_4.
With these support, the inference speed should be faster in the order of Q4_0_8_8 > Q4_0_4_8 > Q4_0_4_4 > Q4_0 without much effect on the quality of response.
This is a list of ARM devices that support different ARM instructions. Apparently, it is only a partial list. It is better you check for i8mm and sve support by yourself.
For Apple devices,
sysctl hw
For other ARM devices (ie most Android devices),
cat /proc/cpuinfo
There are also android apps that can display /proc/cpuinfo.
I was told that for Intel/AMD CPU inference, support for AVX2/AVX512 can also improve the performance of Q4_0_8_8.
On the other hand, Nvidia 3090 inference speed is significantly faster for Q4_0 than the other ggufs. That means for GPU inference, you better off using Q4_0.
Which Q4_0 model to use for ARM devices
| Brand | Series | Model | i8mm | sve | Quant Type |
|---|---|---|---|---|---|
| Apple | A | A4 to A14 | No | No | Q4_0_4_4 |
| Apple | A | A15 to A18 | Yes | No | Q4_0_4_8 |
| Apple | M | M1 | No | No | Q4_0_4_4 |
| Apple | M | M2/M3/M4 | Yes | No | Q4_0_4_8 |
| Tensor | G1,G2 | No | No | Q4_0_4_4 | |
| Tensor | G3,G4 | Yes | Yes | Q4_0_8_8 | |
| Samsung | Exynos | 2200,2400 | Yes | Yes | Q4_0_8_8 |
| Mediatek | Dimensity | 9000 | Yes | Yes | Q4_0_8_8 |
| Mediatek | Dimensity | 9300 | Yes | No | Q4_0_4_8 |
| Qualcomm | Snapdragon | 8 Gen 1 | Yes | Yes | Q4_0_8_8 |
| Qualcomm | Snapdragon | 8 Gen 2,8 Gen 3,X Elite | Yes | No | Q4_0_4_8 |
imatrix quantization
According to this blog, adding imatrix to low bit quant can significantly improve performance. The best dataset for Japanese is MTFMC/imatrix-dataset-for-japanese-llm. Therefore, I also created the imatrix versions of different Q4_0 quants.
However, based on my benchmarking results, the difference is not significant.
Convert safetensors to f16 gguf
Make sure you have llama.cpp git cloned:
python3 convert_hf_to_gguf.py gemma-2-2b-jpn-it/ --outfile gemma-2-2b-jpn-it.f16.gguf --outtype f16
Convert f16 gguf to Q8_0 gguf without imatrix
Make sure you have llama.cpp compiled:
./llama-quantize gemma-2-2b-jpn-it.f16.gguf gemma-2-2b-jpn-it.Q8_0.gguf q8_0
Convert f16 gguf to other ggufs with imatrix
First, prepare imatrix from f16 gguf and c4_en_ja_imatrix.txt
./llama-imatrix -m gemma-2-2b-jpn-it.f16.gguf -f c4_en_ja_imatrix.txt -o gemma-2-2b-jpn-it.imatrix --chunks 32
Then, convert f16 gguf with imatrix to create imatrix gguf
./llama-quantize --imatrix gemma-2-2b-jpn-it.imatrix gemma-2-2b-jpn-it.f16.gguf gemma-2-2b-jpn-it-imatrix.Q4_0_8_8.gguf q4_0_8_8
Downloading using huggingface-cli
First, make sure you have hugginface-cli installed:
pip install -U "huggingface_hub[cli]"
Then, you can target the specific file you want:
huggingface-cli download ymcki/gemma-2-2b-jpn-it-GGUF --include "gemma-2-2b-jpn-it-Q8_0.gguf" --local-dir ./
Credits
Thank you bartowski for providing a README.md to get me started.
Thank you YoutechA320U for the ELYZA-tasks-100 auto evaluation tool.