4 bits quantization of LLaMA using GPTQ
GPTQ is SOTA one-shot weight quantization method
This code is based on GPTQ
Changed to use only pytorch instead of the current cuda kernel. It has no impact on memory usage. There is a slowdown below 128 length, but much faster at 128 and above.
Changed to support new features proposed by GPTQ.
- Slightly adjusted preprocessing of C4 and PTB for more realistic evaluations (used in our updated results); can be activated via the flag --new-eval.
- two new tricks:--act-order (quantizing columns in order of decreasing activation size) and --true-sequential (performing sequential quantization even within a single Transformer block). Those fix GPTQ's strangely bad performance on the 7B model (from 7.15 to 6.09 Wiki2 PPL) and lead to slight improvements on most models/settings in general.
LLaMA-7B(click me)
| LLaMA-7B | Bits | group-size | memory(MiB) | Wikitext2 | checkpoint size(GB) |
|---|---|---|---|---|---|
| FP16 | 16 | - | 13940 | 5.68 | 12.5 |
| RTN | 4 | - | - | 6.29 | - |
| GPTQ | 4 | - | 4740 | 6.09 | 3.5 |
| RTN | 3 | - | - | 25.54 | - |
| GPTQ | 3 | - | 3852 | 8.07 | 2.7 |
| GPTQ | 3 | 128 | 4116 | 6.61 | 3.0 |
LLaMA-13B
| LLaMA-13B | Bits | group-size | memory(MiB) | Wikitext2 | checkpoint size(GB) |
|---|---|---|---|---|---|
| FP16 | 16 | - | OOM | 5.09 | 24.2 |
| RTN | 4 | - | - | 5.53 | - |
| GPTQ | 4 | - | 8410 | 5.36 | 6.5 |
| RTN | 3 | - | - | 11.40 | - |
| GPTQ | 3 | - | 6870 | 6.63 | 5.1 |
| GPTQ | 3 | 128 | 7277 | 5.62 | 5.4 |
LLaMA-33B
| LLaMa-33B | Bits | group-size | memory(MiB) | Wikitext2 | checkpoint size(GB) |
|---|---|---|---|---|---|
| FP16 | 16 | - | OOM | 4.10 | 60.5 |
| RTN | 4 | - | - | 4.54 | - |
| GPTQ | 4 | - | 19493 | 4.45 | 15.7 |
| RTN | 3 | - | - | 14.89 | - |
| GPTQ | 3 | - | 15493 | 5.69 | 12.0 |
| GPTQ | 3 | 128 | 16566 | 4.80 | 13.0 |
LLaMA-65B
| LLaMA-65B | Bits | group-size | memory(MiB) | Wikitext2 | checkpoint size(GB) |
|---|---|---|---|---|---|
| FP16 | 16 | - | OOM | 3.53 | 121.0 |
| RTN | 4 | - | - | 3.92 | - |
| GPTQ | 4 | - | OOM | 3.84 | 31.1 |
| RTN | 3 | - | - | 10.59 | - |
| GPTQ | 3 | - | OOM | 5.04 | 23.6 |
| GPTQ | 3 | 128 | OOM | 4.17 | 25.6 |
Quantization requires a large amount of CPU memory. However, the memory required can be reduced by using swap memory.
Depending on the GPUs/drivers, there may be a difference in performance, which decreases as the model size increases.(IST-DASLab/gptq#1)
According to GPTQ paper, As the size of the model increases, the difference in performance between FP16 and GPTQ decreases.
If you don't have conda, install it first.
conda create --name gptq python=3.9 -y
conda activate gptq
conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia
# Or, if you're having trouble with conda, use pip with python3.9:
# pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
git clone https://github.com/qwopqwop200/GPTQ-for-LLaMa
cd GPTQ-for-LLaMa
pip install -r requirements.txt
torch: tested on v2.0.0+cu117transformers: tested on v4.28.0.dev0datasets: tested on v2.10.1safetensors: tested on v0.3.0- (to run 4-bit kernels: setup for compiling PyTorch CUDA extensions, see also https://pytorch.org/tutorials/advanced/cpp_extension.html, tested on CUDA 11.7)
All experiments were run on a single NVIDIA RTX3090.
#convert LLaMA to hf
python convert_llama_weights_to_hf.py --input_dir /path/to/downloaded/llama/weights --model_size 7B --output_dir ./llama-hf
# Benchmark language generation with 4-bit LLaMA-7B:
# Save compressed model
CUDA_VISIBLE_DEVICES=0 python llama.py ./llama-hf/llama-7b c4 --wbits 4 --true-sequential --act-order --groupsize 128 --save llama7b-4bit-128g.pt
# Or save compressed `.safetensors` model
CUDA_VISIBLE_DEVICES=0 python llama.py ./llama-hf/llama-7b c4 --wbits 4 --true-sequential --act-order --groupsize 128 --save_safetensors llama7b-4bit-128g.safetensors
# Benchmark generating a 2048 token sequence with the saved model
CUDA_VISIBLE_DEVICES=0 python llama.py ./llama-hf/llama-7b c4 --wbits 4 --groupsize 128 --load llama7b-4bit-128g.pt --benchmark 2048 --check
# Benchmark FP16 baseline, note that the model will be split across all listed GPUs
CUDA_VISIBLE_DEVICES=0,1,2,3,4 python llama.py ./llama-hf/llama-7b c4 --benchmark 2048 --check
# model inference with the saved model
CUDA_VISIBLE_DEVICES=0 python llama_inference.py ./llama-hf/llama-7b --wbits 4 --groupsize 128 --load llama7b-4bit-128g.pt --text "this is llama"
# model inference with the saved model with offload(This is very slow. This is a simple implementation and could be improved with technologies like flexgen(https://github.com/FMInference/FlexGen).
CUDA_VISIBLE_DEVICES=0 python llama_inference_offload.py ./llama-hf/llama-7b --wbits 4 --groupsize 128 --load llama7b-4bit-128g.pt --text "this is llama" --pre_layer 16
It takes about 180 seconds to generate 45 tokens(5->50 tokens) on single RTX3090 based on LLaMa-65B. pre_layer is set to 50.
Basically, 4-bit quantization and 128 groupsize are recommended.
This code is based on GPTQ
Thanks to Meta AI for releasing LLaMA, a powerful LLM.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent