4 bits quantization of LLaMa using GPTQ
GPTQ is SOTA one-shot weight quantization method
This code is based on GPTQ
| Model(LLaMa-7B) | Bits | group-size | Wikitext2 | PTB | C4 |
|---|---|---|---|---|---|
| FP16 | 16 | - | 5.67 | 8.79 | 7.05 |
| RTN | 4 | - | 6.28 | 9.68 | 7.70 |
| GPTQ | 4 | - | 6.79 | 10.67 | 8.28 |
| GPTQ | 4 | 64 | 6.16 | 9.66 | 7.52 |
| RTN | 3 | - | 25.66 | 61.25 | 28.19 |
| GPTQ | 3 | - | 20.86 | 37.54 | 22.19 |
| GPTQ | 3 | 64 | 12.24 | 16.77 | 9.55 |
| Model(LLaMa-13B) | Bits | group-size | Wikitext2 | PTB | C4 |
|---|---|---|---|---|---|
| FP16 | 16 | - | 5.08 | 8.06 | 6.58 |
| RTN | 4 | - | 5.52 | 8.62 | 6.96 |
| GPTQ | 4 | - | 5.35 | 8.40 | 6.82 |
| GPTQ | 4 | 64 | 5.18 | 8.18 | 6.66 |
| RTN | 3 | - | 11.41 | 21.21 | 13.20 |
| GPTQ | 3 | - | 6.80 | 10.45 | 8.31 |
| GPTQ | 3 | 64 | 5.50 | 8.60 | 7.00 |
Quantizing the model requires a large amount of CPU memory. For example, quantizing a LLaMa-13b model requires 42gb, and LLaMa-33b requires more memory than 64gb.
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.
torch: tested on v1.12.1+cu113transformers: tested on v4.27.0.dev0(required)datasets: tested on v2.10.1- (to run 4-bit kernels: setup for compiling PyTorch CUDA extensions, see also https://pytorch.org/tutorials/advanced/cpp_extension.html, tested on CUDA 11.3)
All experiments were run on a single NVIDIA RTX3090.
# Compute full precision (FP16) results
CUDA_VISIBLE_DEVICES=0 python llama.py decapoda-research/llama-7b-hf c4
# Run RTN baseline and compute results
CUDA_VISIBLE_DEVICES=0 python llama.py decapoda-research/llama-7b-hf c4 --wbits 4 --nearest
# Run GPTQ and compute results
CUDA_VISIBLE_DEVICES=0 python llama.py decapoda-research/llama-7b-hf c4 --wbits 4 --groupsize 64
To run other LLaMa models replace llama-7b-hf with one of: llama-13b-hf, llama-30b-hf, llama-65b-hf.
See zeroShot/ folder.
# Install kernels
python setup_cuda.py install
# Benchmark performance for FC2 layer of LLaMa-7B
CUDA_VISIBLE_DEVICES=0 python test_kernel.py
# Benchmark language generation with 4-bit LLaMa-7B:
# Save compressed model
CUDA_VISIBLE_DEVICES=0 python llama.py decapoda-research/llama-7b-hf c4 --wbits 4 --save llama7b-4bit.pt
# Benchmark generating a 2048 token sequence with the saved model
CUDA_VISIBLE_DEVICES=0 python llama.py decapoda-research/llama-7b-hf c4 --wbits 4 --load llama7b-4bit.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 decapoda-research/llama-7b-hf c4 --benchmark 2048 --check
# model inference with the saved model
CUDA_VISIBLE_DEVICES=0 python llama_inference.py decapoda-research/llama-7b-hf --wbits 4 --load llama7b-4bit.pt --text "this is llama"
CUDA Kernels support 2,3,4,8 bits.
Basically, 4-bit quantization is recommended.
cuda kernel does not support group size.
| Model | Bits | memory(MiB) | benchmark(ppl) | Wikitext2 | PTB | C4 | checkpoint size(GB) |
|---|---|---|---|---|---|---|---|
| LLaMa-7B with FP16 | 16 | 13940 | 5.23 | 5.67 | 8.79 | 7.05 | 12.5 |
| LLaMa-13B with FP16 | 16 | OOM | - | 5.08 | 8.06 | 6.58 | 24.2 |
| LLaMa-7B with GPTQ | 8 | 7748 | 5.39 | 5.67 | 8.81 | 7.08 | 6.5 |
| LLaMa-13B with GPTQ | 8 | 14570 | 5.00 | 5.09 | 8.06 | 6.61 | 12.4 |
| LLaMa-7B with GPTQ | 4 | 4740 | 6.23 | 6.79 | 10.67 | 8.28 | 3.5 |
| LLaMa-13B with GPTQ | 4 | 8410 | 5.14 | 5.35 | 8.40 | 6.82 | 6.5 |
| LLaMa-7B with GPTQ | 3 | 3852 | 11.43 | 17.94 | 31.44 | 19.65 | 2.75 |
| LLaMa-13B with GPTQ | 3 | 6870 | 5.58 | 6.77 | 10.29 | 8.34 | 5.06 |
| LLaMa-7B with GPTQ | 2 | 3076 | 4152 | 30749 | 45936 | 5045 | 2.0 |
| LLaMa-13B with GPTQ | 2 | 5275 | 6903 | 13203 | 1384 | 8.34 | 5.06 |
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