Sequoia: Scalable, Robust, and Hardware-aware Speculative Decoding

[paper]

Environment Set Up

We recommend the following commands to set up the environment

pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu121
pip install transformers==4.36.2
pip install accelerate==0.26.1
pip install datasets==2.16.1
pip install einops
pip install protobuf
pip install sentencepiece
pip install typing-extensions

Evaluations

To reproduce the main results

cd tests
bash run_L40.sh

or bash run_A100.sh

A command should be in the format like

python testbed.py --model  JackFram/llama-68m   --target meta-llama/Llama-2-7b-hf  \
--T 0.6 --P 1.0  --start 0 --end 200 --M 384 \
--growmap ../A100_growmaps/68m_7b/growmaps/A100-CNN-68m-7b-stochastic.pt \
--Mode greedy --dataset cnn

testbed.py is for stochastic decoding. testbed_greedy.py is for greedy decoding. test_specinfer.py is for specinfer sampling. test_greedyS.py is for Top-k/greedy sampling. test_accept.py is for preparing the accepting rate vector.

--model specifies the draft and --target specifies the target. Currently, only Llama models are supported (including Llama2, Sheared-LLaMA, Vicuna and TinyLlama).

--T specifies the temperature and --P specifies the top-p for generation.

--dataset should be in cnn, openwebtext, c4. --start and --end decides how many examples will be evaluated. --seed is for adjusting random seeds. To precisely reproduce the results, seed is set to be 17 by default.

--growmap specifies the tree structure. We have prepared some growmaps in A100_growmaps and L40_growmaps.

--M should be set at least #tree + 256. 384 is enough for all the experiments except offloading. To run offloading, we need the command like the following

CUDA_VISIBLE_DEVICES=0 python testbed.py --model meta-llama/Llama-2-7b-hf \
--target meta-llama/Llama-2-70b-hf  --T 0.6 --P 1.0 \
--start 0 --end 100 --Mode greedy  --M 1024 \
--growmap  ../L40_growmaps/L40-CNN-7b-70b-stochastic.pt  --offloading --dataset cnn

All experiments in test have the max sequence length of 256. To change this, max_target_seq should be passed to SpecTree. Again, --M should be set at least #tree + max_target_seq.

How to obtain acceptance rate vector

To obtain the acceptance rate vector, which is used in tree_search.py, we need the following command

python test_accept.py --model  JackFram/llama-68m   --target meta-llama/Llama-2-7b-hf  \
--T 0.6 --P 1.0  --start 0 --end 200 --M 288 --W 32\
--ALG stochastic --dataset cnn \

--ALG is stochastic or greedy. --W is the maximum width. --M should be set at least --W + 256.

To statically obtain the acceptance rate vector (which is much faster if the target model needs offloading)

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5 python fast_test.py --model meta-llama/Llama-2-7b-hf  \
--target meta-llama/Llama-2-70b-hf --T 1.1 --P 1.0 --DP 1.1 --W 32 --start 0 --end 200

The acceptance rate vector will be printed and will be saved to --dst (../acceptance-rate-vector.pt by default).

How to generate growmaps

We use the following command

python tree_search.py --config demo-config.json

We can modify the content of demo-config.json to generate different growmaps. The growmaps for experiments in the paper in prepared in L40_growmaps and A100_growmaps.

TODOs

• Support other open source models.
• Support multi-round dialogue.
• Support INT4/8 quantization.
• Support multi-GPU.

Citation

If you find Sequoia useful or relevant to your project and research, please kindly cite our paper:

@article{chen2024sequoia,
  title={Sequoia: Scalable, Robust, and Hardware-aware Speculative Decoding},
  author={Chen, Zhuoming and May, Avner and Svirschevski, Ruslan and Huang, Yuhsun and Ryabinin, Max and Jia, Zhihao and Chen, Beidi},
  journal={arXiv preprint arXiv:2402.12374},
  year={2024}
}