| Blog |
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. Theoretically driven (stay tuned for our upcoming paper), EAGLE is building upon the following First Principle:
The sequence of LLM feature vectors is compressible over time, making the prediction of subsequent feature vectors from previous ones easy.
- EAGLE is:
- 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.
Inference is conducted on RTX 3090 GPUs at fp16 precision using the Vicuna 33B model. For an enhanced viewing experience, the animation has been sped up fourfold.
<<<<<<< HEAD
- Support non-greedy inference (provably maintaining text distribution).
- Support bs > 1.
- Support vLLM.
- Support more LLMs such as Mixtral 8x7B.
=======
2e661b5 (add: temp)
2e661b5 (add: temp)
pip install -r requirements.txt
Base Model | EAGLE on Hugging Face | # EAGLE Parameters | Base Model | EAGLE on Hugging Face | # EAGLE Parameters |
---|---|---|---|---|---|
Vicuna 7B | yuhuili/EAGLE-Vicuna-7B-v1.3 | 0.24B | LLaMA2-Chat 7B | yuhuili/EAGLE-llama2-chat-7B | 0.24B |
Vicuna 13B | yuhuili/EAGLE-Vicuna-13B-v1.3 | 0.37B | LLaMA2-Chat 13B | yuhuili/EAGLE-llama2-chat-13B | 0.37B |
Vicuna 33B | yuhuili/EAGLE-Vicuna-33B-v1.3 | 0.56B | LLaMA2-Chat 70B | yuhuili/EAGLE-llama2-chat-70B | 0.99B |
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 application.webui --ea-model-path [path of EAGLE weight]\
--base-model-path [path of the original model]\
--model-type [vicuna or llama-2-chat]
<<<<<<< HEAD
=======
2e661b5 (add: temp) You can use our provided "eagenerate" for speedup generation just like using 'generate' from Hugging Face. Here is an example.
from 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"
)
model.eval()
your_message="Hello"
if ues_llama_2_chat:
conv = get_conversation_template("llama-2-chat")
sys_p = "You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.\n\nIf a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information."
conv.system_message = sys_p
conv.append_message(conv.roles[0], your_message)
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt() + " "
if use_vicuna:
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 and LLaMA2-Chat 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. The current repository only supports a batch size of 1. We plan to update it in the future to support a batch size greater than 1.
You can run the following command to generate the training data.
python -m ge_data.allocation --outdir [path of data]
cd train
accelerate launch --mixed_precision=bf16 main.py --tmpdir [path of data]\
--cpdir [path of checkpoints]
You can test the speed of EAGLE on MT-bench using the following command.
python -m 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 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.
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, and others.