This is GreenBitAI's research code for running 2-bit and 1-bit LLaMA models with extreme compression yet still strong performance, the quantized models are available on the model zoo.
This is meant to be a research demo for the quality of the model. There is no speed-up implemented yet.
Over the next few months, we will continue offering 2-bit and 1-bit versions of LLaMA models. Additionally, we are considering the provision of low-bit versions for other open-source LLMs in the future.
[12/14/2023] We are happy to release the lossless (<1%) W4A16 01-Yi models (low_bit_yi branch). The 2-bit version will be made open soon.
[10/04/2023] We are happy to release the W2A16 g8/32 TinyLLaMA-1.1B models.
[09/29/2023] We are happy to release the W2A16 g8 LLaMA-1 30B and LLaMA-2 70B models.
[09/12/2023] We are happy to announce the release of the 2-bit LLaMA-2 7B (W2A16 g32/g8) models.
[08/31/2023] We are happy to release the harness benchmarks on 14 zero-shot tasks based on our 2-bit models. Happy trying ππ.
[08/16/2023] We are happy to release the 2-bit OpenLLaMA 3B models, which are quantized into 2-bit representation yet still with strong performance πβ.
LLM Models | Method | Bits | Groupsize | Wikitext2 | C4 | Checkpoint Size (GiB) |
---|---|---|---|---|---|---|
LLaMA-2-70B1 | FP16 | 16 | - | 3.31 | 5.70 | 130 |
Ours | 2 | 8 | 3.87 | 5.96 | 26.9 | |
LLaMA-1-30B1 | FP16 | 16 | - | 4.10 | 5.98 | 60.5 |
Ours | 2 | 8 | 4.75 | 6.57 | 12.9 | |
LLaMA-2-7B1 | FP16 | 16 | - | 5.47 | 6.97 | 12.5 |
GPTQ2 | 4 | 128 | 5.61 | 7.12 | 3.6 | |
GPTQ2 | 2 | 128 | 2.2e5 | 1.7e5 | 2.2 | |
OmniQuant3 | 4 | 128 | 5.58 | 7.12 | 3.8 | |
OmniQuant3 | 3 | 128 | 6.03 | 7.35 | 3.2 | |
OmniQuant3 | 2 | 128 | 12.84 | 17.40 | 2.2 | |
OmniQuant3 | 2 | 64 | 10.56 | 13.77 | - | |
Ours | 4 | 32 | 5.55 | 7.08 | 3.7 | |
Ours | 2 | 8 | 6.09 | 7.63 | 2.9 | |
Ours | 2 | 32 | 7.13 | 8.67 | 2.2 | |
LLaMA-1-7B4 | FP16 | 16 | - | 5.67 | 7.07 | 12.5 |
GPTQ2 | 4 | 128 | 5.85 | 7.21 | 3.6 | |
GPTQ2 | 3 | 128 | 6.61 | 7.85 | 3.0 | |
OmniQuant3 | 2 | 128 | 10.53 | 13.89 | 2.2 | |
Ours | 2 | 32 | 7.59 | 8.96 | 2.2 | |
LLaMA 3B5 | FP16 | 16 | - | 7.34 | 9.33 | 6.8 |
GPTQ2 | 4 | 128 | 7.54 | 9.58 | 1.9 | |
Ours | 4 | 32 | 7.43 | 9.51 | 2.0 | |
Ours | 2 | 8 | 8.32 | 10.56 | 1.5 | |
Ours | 2 | 16 | 8.92 | 11.29 | 1.3 | |
Ours | 2 | 32 | 9.82 | 12.14 | 1.2 | |
TinyLLaMA 1.1B6 | FP16 | 16 | - | 9.10 | 10.6 | 4.0 |
Ours | 2 | 8 | 9.99 | 11.75 | 0.6 | |
Ours | 2 | 32 | 12.04 | 14.27 | 0.5 |
LLM Models | Method | Bits | Checkpoint Size (GiB) |
---|---|---|---|
LLaMA-2-70B-Chat1 | FP16 | 16 | 130 |
Ours | 2 | 26.9 | |
CodeLLaMA-34B7 | FP16 | 16 | 63 |
Ours | 2 | 13.5 | |
CodeLLaMA-34B-Python7 | FP16 | 16 | 63 |
Ours | 2 | 13.5 | |
CodeLLaMA-34B-Instruction7 | FP16 | 16 | 63 |
Ours | - | - |
Task | Metric | TinyLLaMA 1.1B q2g32 | TinyLLaMA 1.1B q2g8 | LLaMA 3B q2g32 | LLaMA 3B q2g16 | LLaMA 3B q2g8 | LLaMA-1 7B q2g32 | LLaMA-2 7B q2g32 | LLaMA-2 7B q2g8 | LLaMA 1.1B FP16 | LLaMA 3B FP16 | LLaMA-1 7B FP16 |
---|---|---|---|---|---|---|---|---|---|---|---|---|
Openbookqa | acc | 0.152 | 0.192 | 0.196 | 0.238 | 0.242 | 0.224 | 0.246 | 0.296 | 0.208 | 0.27 | 0.29 |
ac_norm | 0.328 | 0.338 | 0.332 | 0.358 | 0.362 | 0.388 | 0.376 | 0.4 | 0.368 | 0.4 | 0.41 | |
arc_challenge | acc | 0.3268 | 0.2278 | 0.279 | 0.2978 | 0.3148 | 0.3422 | 0.3268 | 0.3618 | 0.243 | 0.34 | 0.39 |
ac_norm | 0.3387 | 0.273 | 0.2944 | 0.3319 | 0.3345 | 0.3387 | 0.3387 | 0.372 | 0.288 | 0.37 | 0.41 | |
hellawswag | acc | 0.34 | 0.3769 | 0.4238 | 0.444 | 0.462 | 0.4996 | 0.4961 | 0.5379 | 0.403 | 0.49 | 0.68 |
ac_norm | 0.4097 | 0.4711 | 0.5685 | 0.5988 | 0.6242 | 0.6447 | 0.6464 | 0.7014 | 0.503 | 0.67 | 0.73 | |
piqa | acc | 0.6518 | 0.6931 | 0.7024 | 0.716 | 0.7291 | 0.7476 | 0.7503 | 0.7715 | 0.71 | 0.75 | 0.78 |
ac_norm | 0.6393 | 0.6812 | 0.7116 | 0.7247 | 0.7312 | 0.7443 | 0.7421 | 0.7568 | 0.688 | 0.76 | 0.78 | |
arc_easy | acc | 0.4411 | 0.5109 | 0.5997 | 0.646 | 0.6528 | 0.6061 | 0.6174 | 0.6254 | 0.533 | 0.69 | 0.68 |
ac_norm | 0.3716 | 0.412 | 0.5417 | 0.58 | 0.5972 | 0.4566 | 0.4781 | 0.4958 | 0.43 | 0.65 | 0.52 | |
Winogrande | acc | 0.532 | 0.5249 | 0.5683 | 0.5888 | 0.6054 | 0.6283 | 0.6298 | 0.6582 | 0.558 | 0.62 | 0.68 |
boolq | acc | 0.592 | 0.6174 | 0.6281 | 0.6636 | 0.6327 | 0.6425 | 0.7061 | 0.7242 | 0.583 | 0.68 | 0.75 |
truthfulqa_mc | mc1 | 0.2338 | 0.2277 | 0.2509 | 0.2118 | 0.2252 | 0.224 | 0.2313 | 0.2399 | 0.228 | 0.22 | 0.21 |
mc2 | 0.4211 | 0.406 | 0.3962 | 0.3501 | 0.3625 | 0.3702 | 0.3854 | 0.3795 | 0.401 | 0.35 | 0.34 | |
anli_r1 | acc | 0.363 | 0.336 | 0.337 | 0.334 | 0.344 | 0.331 | 0.333 | 0.363 | 0.354 | 0.33 | 0.35 |
anli_r2 | acc | 0.331 | 0.346 | 0.335 | 0.332 | 0.331 | 0.326 | 0.349 | 0.347 | 0.341 | 0.32 | 0.34 |
anli_r3 | acc | 0.3758 | 0.3633 | 0.3358 | 0.3383 | 0.3425 | 0.3417 | 0.36 | 0.3733 | 0.358 | 0.35 | 0.37 |
wic | acc | 0.5 | 0.5 | 0.4984 | 0.5094 | 0.4969 | 0.4984 | 0.4953 | 0.489 | 0.5 | 0.48 | 0.5 |
rte | acc | 0.4874 | 0.4874 | 0.5596 | 0.5993 | 0.5632 | 0.639 | 0.6065 | 0.6426 | 0.516 | 0.58 | 0.56 |
record | f1 | 0.7608 | 0.8023 | 0.8502 | 0.8625 | 0.8687 | 0.8859 | 0.8872 | 0.9037 | 0.82 | 0.88 | 0.91 |
em | 0.753 | 0.7934 | 0.8427 | 0.8545 | 0.8612 | 0.8781 | 0.8801 | 0.8959 | 0.818 | 0.89 | 0.91 | |
Average | 0.438 | 0.4498 | 0.4881 | 0.5037 | 0.5087 | 0.5122 | 0.5181 | 0.5391 | 0.469 | 0.528 | 0.5519 | |
model size | GiB | 0.5 | 0.6 | 1.2 | 1.3 | 1.5 | 2.2 | 2.2 | 2.9 | 4.4 | 6.8 | 12.5 |
The inference currently requires a machine with CUDA installed. Then you can simply run:
pip install -r requirements.txt
Use the environment variable CUDA_VISIBLE_DEVICES
to select the correct GPU.
Multi-GPU is not supported, but the model is very compressed, so 1 GPU should be enough.
To use the instruction-tuned model, you can use the following commands
in scripts/
. Predefined scripts already there:
bash scripts/evaluate/tiny_llama_w2a16g32.sh # for open task evaluation of the base model.
bash scripts/inference/llama2_70b_w2a16g8.sh # for text generation inference of the base model.
bash scripts/instruction-chat/llama2_70b_w2a16g8.sh # for instruction following chat of the fine-tuned model.
bash scripts/inference/codellama_34b_w2a16g8.sh # for text generation inference of the codellama model
This code is based on:
Thanks to Meta AI for releasing LLaMA, a powerful LLM.
If you use our approach in your research, please cite our work as follows:
@article{low_bit_llama,
title={Advanced Ultra-Low Bitrate Compression Techniques for the LLaMA Family of LLMs},
author={Guo, Nianhui and Bethge, Joseph and Hu, Ting and Meinel, Christoph and Yang, Haojin},
journal={https://github.com/GreenBitAI/low_bit_llama},
year={2023}
}
The original code was released under its respective license and copyrights, i.e.:
datautils.py
andevaluate.py
: GPTQ for LLaMA released under Apache 2.0 Licensemodel.py
,peft_tuners_lora.py
andinference.py
(basis forllama_2b_*.py
files): Alpaca_lora_4bit released under MIT License
We release our changes and additions to these files under the Apache 2.0 License.