/EAGLE

EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty

Primary LanguagePythonApache License 2.0Apache-2.0

EAGLE

 EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty

| Paper | Blog |

Version License Maintenance Contributions welcome

benchmark

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 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:
    • achieving 2x speedup on gpt-fast, one of the fastest-known open-sourced inferences.
    • 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.

demogif

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.

Update

2023.12.8: EAGLE v1.0 is released.

2024.1.15: We now support batch size > 1 generation.

2024.1.17: We have integrated gpt-fast into EAGLE, further accelerating the generation speed.

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

Todo

  • Support non-greedy inference (provably maintaining text distribution).
  • Support bs > 1.
  • Support gpt-fast.
  • Support more LLMs such as Mixtral 8x7B.
  • Support vLLM.

Contents

Setup & Installation

pip install -r requirements.txt

EAGLE Weights

Base Model EAGLE on Hugging Face # EAGLE Parameters Base Model EAGLE on Hugging Face # EAGLE Parameters
Vicuna-7B-v1.3 yuhuili/EAGLE-Vicuna-7B-v1.3 0.24B LLaMA2-Chat 7B yuhuili/EAGLE-llama2-chat-7B 0.24B
Vicuna-13B-v1.3 yuhuili/EAGLE-Vicuna-13B-v1.3 0.37B LLaMA2-Chat 13B yuhuili/EAGLE-llama2-chat-13B 0.37B
Vicuna-33B-v1.3 yuhuili/EAGLE-Vicuna-33B-v1.3 0.56B LLaMA2-Chat 70B yuhuili/EAGLE-llama2-chat-70B 0.99B
Mixtral-8x7B-Instruct-v0.1 yuhuili/EAGLE-mixtral-instruct-8x7B 0.28B

Inference

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.

With UI

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]

With Code

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.

Batch size > 1

Switch to the bsne1 branch.

git checkout bsne1

Here is an example. Note that left padding is needed.

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"
)
# left padding
model.eval()
model.tokenizer.padding_side = "left"
model.tokenizer.pad_token = model.tokenizer.eos_token
model.config.pad_token_id = model.config.eos_token_id

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."

your_message="Compose an engaging travel blog post about a recent trip to Hawaii, highlighting cultural experiences and must-see attractions."
conv = get_conversation_template("llama-2-chat")
conv.system_message = sys_p
conv.append_message(conv.roles[0], your_message)
conv.append_message(conv.roles[1], None)
prompt1 = conv.get_prompt()+" "

your_message="Hello"
conv = get_conversation_template("llama-2-chat")
conv.system_message = sys_p
conv.append_message(conv.roles[0], your_message)
conv.append_message(conv.roles[1], None)
prompt2 = conv.get_prompt()+" "

input_s=model.tokenizer([prompt1,prompt2],return_tensors="pt",padding=True).to("cuda")
output_ids=model.eagenerate(input_s.input_ids,input_s.attention_mask,temperature=0.0,max_new_tokens=512,top_k=15)
output=model.tokenizer.batch_decode(output_ids)
print(output)

# vanilla auto-regression
# output_ids, new_token, idx=model.naivegenerate(input_s.input_ids,input_s.attention_mask,temperature=0.0,max_new_tokens=512,top_k=15,log=True)

Train

Generate Train Data

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

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

Train the Auto-regression Head

cd train
accelerate launch --mixed_precision=bf16 main.py --tmpdir [path of data]\
--cpdir [path of checkpoints]

Evaluation

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.

With gpt-fast

GPT-Fast primarily accelerates generation through quantization and compilation, which we have integrated into EAGLE. Here is the result of an experiment conducted on MT-bench with a single RTX3090, using LLaMA2-chat 7B.

Precision fp16 int4
vanilla 24.5 tokens/s N/A
gpt-fast 55.1 tokens/s 106.9 tokens/s
EAGLE+gpt-fast 100.2 tokens/s 160.4 tokens/s

demogif

Inference is conducted on a single RTX3090 GPU at int4 precision using the LLaMA2-chat 7B model. No additional training required.

In EAGLE, using gpt-fast only requires three steps: setting up the environment, quantizing weights, and modifying the model path.

Setup

Switch to the eaglefast branch.

git checkout eaglefast

Install the Preview (Nightly) version of PyTorch with CUDA 12.1, do not use "pip install torch" as it installs the Stable version, which lacks some of the new features used by gpt-fast.

This is a requirement for gpt-fast, whereas other branches of eagle can use the Stable version of PyTorch.

Quantizing Weights

Convert Huggingface weights to the format required by gpt-fast.

python convert/convert_hf_checkpoint.py --checkpoint_dir path_of_base_model
python convert/convert_hf_checkpoint_EAGLE.py --checkpoint_dir path_of_eagle

Quantize weights.

python -m model.quantize_llama --checkpoint_path path_of_base_model/model.pth
python -m model.quantize_EAGLE --checkpoint_path path_of_eagle/model.pth

Modifying Path

When specifying the model weights (including the base model and EAGLE), change "path" to "path/model_int4.g32.pth".

🌟 Our Contributors

A heartfelt thank you to all our contributors.

Contributors

Reference

For technical details and full experimental results, please check the paper.

@article{li2024eagle, 
	author = {Yuhui Li and Fangyun Wei and Chao Zhang and Hongyang Zhang}, 
	title = {EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty}, 
	journal = {arXiv preprint arXiv:2401.15077},
	year = {2024}
}

Acknowledgements

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.