File size: 3,337 Bytes
2b6d729 8b8d414 d2df56b 0911327 d2df56b 0911327 8b8d414 d2df56b 740eb40 cc8d49b d2df56b 69d5187 0911327 69d5187 3c37a9f 69d5187 d0dd825 740eb40 69d5187 d2df56b 62814d1 69d5187 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 |
---
license: apache-2.0
datasets:
- LLM360/AmberDatasets
language:
- en
metrics:
- accuracy
pipeline_tag: text-generation
---
# FBI-LLM-7B
## Model Details
This work presents a Fully BInarized Large Language Model (FBI-LLM), demonstrating for the first time how to train a large-scale binary language model (not the ternary LLM like BitNet b1.58 from scratch to match the performance of its full-precision counterparts (e.g., FP16 or BF16) in transformer-based LLMs. It achieves this by employing an autoregressive distillation (AD) loss with maintaining equivalent model dimensions (130M, 1.3B, 7B) and training data volume as regular LLM pretraining, while delivering competitive results in terms of perplexity and task-specific effectiveness. Intriguingly, by analyzing the training trajectory, we find that the pretrained weight is not necessary for training binarized LLMs from scratch. This research encourages a new computational framework and may facilitate the future design of specialized hardware tailored for fully 1-bit LLMs. We make all models, code, and training dataset fully accessible and transparent to support further research.
**Input**: Models input text only.
**Output**: Models generate text only.
## Tokenizer
We use the same tokenizer as [meta-llama/Llama-2-7b-hf](https://huggingface.co/meta-llama/Llama-2-7b-hf)
## Training Data
We use [AmberDateset](https://huggingface.co/datasets/LLM360/AmberDatasets) to train our models.
## Result
<img src=https://huggingface.co/LiqunMa/FBI-LLM_7B/resolve/main/main_result.jpg width="90%" />
<!--  -->
## How to use
Please download the code from [LiqunMa/FBI-LLM](https://github.com/LiqunMa/FBI-LLM) firstly
```python
from pathlib import Path
from transformers import AutoTokenizer,LlamaConfig,LlamaForCausalLM, AutoModelForCausalLM
from qat.replace_module import replace_with_learnable_binarylinear
def load_model(model_size, model_dir):
assert model_size in ["130M", "1.3B", "7B"]
model_dir = Path(model_dir)
with Path(f'FBI-LLM_configs/FBI-LLM_llama2_{model_size}.json').open('r') as r_f:
config = json.load(r_f)
llama_config = LlamaConfig(**config)
model = LlamaForCausalLM(llama_config).to('cuda')
tokenizer = AutoTokenizer.from_pretrained('meta-llama/Llama-2-7b-hf', padding_side="right", use_fast=False)
model = replace_with_learnable_binarylinear(model, scaling_pattern = "column", keep_parts = ["lm_head"])
weight_dict = {}
ckpt_plist = [p for p in model_dir.iterdir() if p.suffix == '.bin']
for p in ckpt_plist:
weight_dict = torch.load(p)
for k,v in _weight_dict.items():
if 'self_attn.rotary_emb.inv_freq' not in k:
weight_dict[k] = v
model.load_state_dict(weight_dict)
for param in model.parameters():
param.data = param.data.to(torch.float16)
return model, tokenizer
```
## Citation
### BibTeX:
```
@misc{ma2024fbillmscalingfullybinarized,
title={FBI-LLM: Scaling Up Fully Binarized LLMs from Scratch via Autoregressive Distillation},
author={Liqun Ma and Mingjie Sun and Zhiqiang Shen},
year={2024},
eprint={2407.07093},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2407.07093},
}
``` |