I am currently focusing on AutoGPTQ and recommend using AutoGPTQ instead of GPTQ for Llama.
4 bits quantization of LLaMA using GPTQ
GPTQ is SOTA one-shot weight quantization method
It can be used universally, but it is not the fastest and only supports linux.
Triton only supports Linux, so if you are a Windows user, please use WSL2.
AutoGPTQ-triton, a packaged version of GPTQ with triton, has been integrated into AutoGPTQ.
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 |
| GPTQ | 4 | 128 | 4891 | 5.85 | 3.6 |
| 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 |
| GPTQ | 4 | 128 | 8747 | 5.20 | 6.7 |
| 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 |
| GPTQ | 4 | 128 | 20570 | 4.23 | 16.3 |
| 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 |
| GPTQ | 4 | 128 | OOM | 3.65 | 32.3 |
| 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.
GPTQ vs bitsandbytes
LLaMA-7B(click me)
| LLaMA-7B(seqlen=2048) | Bits Per Weight(BPW) | memory(MiB) | c4(ppl) |
|---|---|---|---|
| FP16 | 16 | 13948 | 5.22 |
| GPTQ-128g | 4.15 | 4781 | 5.30 |
| nf4-double_quant | 4.127 | 4804 | 5.30 |
| nf4 | 4.5 | 5102 | 5.30 |
| fp4 | 4.5 | 5102 | 5.33 |
LLaMA-13B
| LLaMA-13B(seqlen=2048) | Bits Per Weight(BPW) | memory(MiB) | c4(ppl) |
|---|---|---|---|
| FP16 | 16 | OOM | - |
| GPTQ-128g | 4.15 | 8589 | 5.02 |
| nf4-double_quant | 4.127 | 8581 | 5.04 |
| nf4 | 4.5 | 9170 | 5.04 |
| fp4 | 4.5 | 9170 | 5.11 |
LLaMA-33B
| LLaMA-33B(seqlen=1024) | Bits Per Weight(BPW) | memory(MiB) | c4(ppl) |
|---|---|---|---|
| FP16 | 16 | OOM | - |
| GPTQ-128g | 4.15 | 18441 | 3.71 |
| nf4-double_quant | 4.127 | 18313 | 3.76 |
| nf4 | 4.5 | 19729 | 3.75 |
| fp4 | 4.5 | 19729 | 3.75 |
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
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
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 ${MODEL_DIR} 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 ${MODEL_DIR} 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 ${MODEL_DIR} 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 ${MODEL_DIR} c4 --benchmark 2048 --check
# model inference with the saved model
CUDA_VISIBLE_DEVICES=0 python llama_inference.py ${MODEL_DIR} --wbits 4 --groupsize 128 --load llama7b-4bit-128g.pt --text "this is llama"
# model inference with the saved model using safetensors loaded direct to gpu
CUDA_VISIBLE_DEVICES=0 python llama_inference.py ${MODEL_DIR} --wbits 4 --groupsize 128 --load llama7b-4bit-128g.safetensors --text "this is llama" --device=0
# model inference with the saved model with offload(This is very slow).
CUDA_VISIBLE_DEVICES=0 python llama_inference_offload.py ${MODEL_DIR} --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.
You can also export quantization parameters with toml+numpy format.
CUDA_VISIBLE_DEVICES=0 python llama.py ${MODEL_DIR} c4 --wbits 4 --true-sequential --act-order --groupsize 128 --quant-directory ${TOML_DIR}
This code is based on GPTQ
Thanks to Meta AI for releasing LLaMA, a powerful LLM.
Triton GPTQ kernel code is based on GPTQ-triton