| Paper (EAGLE) | Paper (EAGLE-2) | Blog | Demo |

EAGLE (Extrapolation Algorithm for Greater Language-model Efficiency) is a new baseline for fast decoding of Large Language Models (LLMs) with provable performance maintenance. This approach involves extrapolating the second-top-layer contextual feature vectors of LLMs, enabling a significant boost in generation efficiency.

• EAGLE is:
  • certified by the third-party evaluation as the fastest speculative method so far.
  • achieving 2x speedup on gpt-fast.
  • 3x faster than vanilla decoding (13B).
  • 2x faster than Lookahead (13B).
  • 1.6x faster than Medusa (13B).
  • provably maintaining the consistency with vanilla decoding in the distribution of generated texts.
  • trainable (within 1-2 days) and testable on 8x RTX 3090 GPUs. So even the GPU poor can afford it.
  • combinable with other parallelled techniques such as vLLM, DeepSpeed, Mamba, FlashAttention, quantization, and hardware optimization.

EAGLE-2 uses the confidence scores from the draft model to approximate acceptance rates, dynamically adjusting the draft tree structure, which further enhances performance.

• EAGLE-2 is:
  • 4x faster than vanilla decoding (13B).
  • 1.4x faster than EAGLE-1 (13B).

Using EAGLE-2, the inference speed on 2 RTX 3060 GPUs can be faster than vanilla autoregressive decoding on an A100 GPU.

2024.6.27: EAGLE-2 is released.

2024.2.25: EAGLE is certified by the third-party evaluation as the fastest speculative method.

2024.1.17: We now support Mixtral-8x7B-Instruct.

2023.12.8: EAGLE v1.0 is released.

• Support non-greedy inference (provably maintaining text distribution).
• Support more LLMs such as Mixtral 8x7B.
• Support LLaMA-3.
• Support Qwen-2.

• Setup & Installation
• EAGLE Weights
• Inference
  • With UI
  • With Code
• Train
  • Generate Train Data
  • Train the Auto-regression Head
  • Inference on custom models
• Evaluation

git clone https://github.com/SafeAILab/EAGLE.git
cd EAGLE
pip install -r requirements.txt

Compared to EAGLE, EAGLE-2 does not require additional training and uses the same weights.

The inference code we provide automatically allocates model weights (loading a model across multiple GPUs), allowing you to run models that exceed the memory of a single GPU.

We have provided a suggested web interface, which you can use by running the following command. After the model is fully loaded, a URL will be output in the terminal, which you can enter into your browser to access.

python -m eagle.application.webui --ea-model-path [path of EAGLE weight]\ 
		--base-model-path [path of the original model]\
		--model-type [vicuna\llama2\llama3]\
        --total-token [int]

The total-token is the number of draft tokens. For smaller models and advanced GPUs, this value can be set larger. Adjusting according to the specific device and model can achieve better results. If set to -1, EAGLE-2 will automatically configure this parameter.

You can use our provided "eagenerate" for speedup generation just like using 'generate' from Hugging Face. Here is an example.

from eagle.model.ea_model import EaModel
from fastchat.model import get_conversation_template
model = EaModel.from_pretrained(
    base_model_path=base_model_path,
    ea_model_path=EAGLE_model_path,
    torch_dtype=torch.float16,
    low_cpu_mem_usage=True,
    device_map="auto",
    total_token=-1
)
model.eval()
your_message="Hello"
conv = get_conversation_template("vicuna")
conv.append_message(conv.roles[0], your_message)
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt()
input_ids=model.tokenizer([prompt]).input_ids
input_ids = torch.as_tensor(input_ids).cuda()
output_ids=model.eagenerate(input_ids,temperature=0.5,max_new_tokens=512)
output=model.tokenizer.decode(output_ids[0])

Note: Vicuna, LLaMA2-Chat, and LLaMA3-Instruct are both chat models. You need to use the correct chat template, otherwise it will cause abnormal output from the model and affect the performance of EAGLE.

You can run the following command to generate the training data.

python -m eagle.ge_data.allocation --outdir [path of data]

accelerate launch -m --mixed_precision=bf16 eagle.train.main --tmpdir [path of data]\
--cpdir [path of checkpoints] -- configpath [path of config file]

eagle/train provides examples of configuration files.

You can also use DeepSpeed for training.

cd eagle/train
deepspeed main_deepspeed.py --deepspeed_config ds_config.json

If the original LLM structure differs from LLaMA and Mixtral, you can utilize EAGLE as follows:

Copy the modeling_basemodelname.py from the Transformers library and proceed to make modifications to leverage the pre-allocated kv_cache for enhanced speed in the base model. You can refer to model/modeling_llama_kv.py for guidance, where places that require modifications are annotated with # [MODIFIED]. These modifications are minimal.

You can test the speed of EAGLE on MT-bench using the following command.

python -m eagle.evaluation.gen_ea_answer_vicuna(or gen_ea_answer_vicuna_llama2chat)\
		 --ea-model-path [path of EAGLE weight]\ 
		 --base-model-path [path of the original model]\

If you need specific acceleration ratios, you will also need to run the following command to get the speed of vanilla auto-regression.

python -m eagle.evaluation.gen_baseline_answer_vicuna\
		(or gen_ea_answer_vicuna_llama2chat)\
		 --ea-model-path [path of EAGLE weight]\ 
		 --base-model-path [path of the original model]\

The above two commands will each generate a .jsonl file that records the generation results and wall time. Then, you can use evaluation/speed.py to calculate the ratio of speeds.

A heartfelt thank you to all our contributors.

For technical details and full experimental results, please check the paper of EAGLE and the paper of EAGLE-2.

@inproceedings{li2024eagle, 
	author = {Yuhui Li and Fangyun Wei and Chao Zhang and Hongyang Zhang}, 
	title = {EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty}, 
	booktitle = {International Conference on Machine Learning},
	year = {2024}
}
@misc{li2024eagle2fasterinferencelanguage,
      title={EAGLE-2: Faster Inference of Language Models with Dynamic Draft Trees}, 
      author={Yuhui Li and Fangyun Wei and Chao Zhang and Hongyang Zhang},
      year={2024},
      eprint={2406.16858},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2406.16858}, 
}

This project has been influenced by many excellent projects in the LLM community, such as Medusa, FastChat, and others. The logo is designed by GPT-4. We also appreciate many valuable discussions with Tianle Cai, Hao Zhang, Ziteng Sun, and others.