Pyramid KV

News

[2024-06-25] Support multi-GPUs inference with big LLMs now! Try out PyramidKV on LlaMa-3-70B-Instruct!
[2024-06-10] Support PyramidKV, SnapKV, H2O and StreamingLLM at Flash Attention v2, Sdpa Attention now! If your devices (i.e., V100, 3090) does not support Flash Attention v2, you can set attn_implementation=sdpa to try PyramidKV at Sdpa Attention!

TODO:

[x] Support implementation of Streaming LLM, H2O and SnapKV
[x] Support Mistral model
[x] Support implementation of Needle
[x] Support KV cache compression without Flash Attention v2 (i.e. Sdpa Attention) for V100
[x] Support multi-GPU inference for 70B LlaMa-3
[ ] Introduce new functions to support kv cache budget allocation (i.e., supports for percentage.)
[ ] Support Mixtral
[ ] Support Batch Inference
[ ] Support KV cache compression at decoding stage

Performence

Visualization: Inefficient Attention

The Llama model attention map with 3 documents is represented as follows:

we provide a notebook visualization.ipynb to reproduce the visualization result of each Llama-2-7b-hf model layer for a given 3 document.

Model attention maps for different layers would be stored at ./attention

Requirements

transformers >= 4.41
flash-attn >= 2.4.0.post1

Installation


git clone https://github.com/Zefan-Cai/PyramidKV.git
cd PyramidKV
pip install -r requirements.txt .

Inference

We support inference code on LongBench to repuduce our result.

Please refer to scripts/scripts_longBench/eval.sh to modify the parameters according to your requirements.

Our codebase support Flash Attention v2, Sdpa Attention, etc. The results presented in our paper in based on Flash Attention v2.

export CUDA_VISIBLE_DEVICES=$1

method=$2 # Support PyramidKV, SnapKV, H2O, StreamingLLM
max_capacity_prompts=64 # 128,2048 in paper
attn_implementation=$3 # Support "flash_attention_2", "sdpa", "eager".
source_path=$4
model_path=$5
save_dir=${source_path}"results_long_bench" # path to result save_dir

python3 run_longbench.py \
    --method ${method} \
    --model_path ${model_path} \
    --max_capacity_prompts ${max_capacity_prompts} \
    --attn_implementation ${attn_implementation} \
    --save_dir ${save_dir} \
    --use_cache True

CUDA_VISIBLE_DEVICES: For multi-GPU inference for big LLMs, just need to specify CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7. For single GPU inference, just need to specify CUDA_VISIBLE_DEVICES=0.
model_path: Path to your model. Support "Llama-3-8B-Instruct" for now.
method: Support PyramidKV, SnapKV, StreamingLLM, H2O.
max_capacity_prompts: Selected KV Size in each layer. （e.g. 128, 2048 in paper）. When method is "PyramidKV", given that the total number of KV remains unchanged, the specific KV length for each layer will be modified accordingly
save_dir: Path to your dir to save LongBench result.

After modifying parameters, run:


sh scripts/scripts_longBench/eval.sh

Needle in haystack

We support inference code on Needle in haystack to repuduce our result.

Please refer to scripts/scripts_needle/eval.sh to modify the parameters according to your requirements.

Our codebase support Flash Attention v2, Sdpa Attention, etc. The results presented in our paper in based on Flash Attention v2.


METHOD='pyramidkv'       # ['full', 'pyramidkv', 'snapkv', 'streamingllm', 'h2o']
MAX_CAPACITY_PROMPT=96  # [64, 96, 128, 256, 512, 1024, 2048, ...]
attn_implementation="flash_attention_2" # Support "flash_attention_2", "sdpa", "".
TAG=test

# For Llama3-8b

(
python -u run_needle_in_haystack.py --s_len 1000 --e_len 8001\
    --model_provider LLaMA3 \
    --model_name /mnt/workspace/zhiyuanhu/yuliang/models/llama3-8b_raw \
    --attn_implementation ${attn_implementation} \
    --step 100 \
    --method $METHOD \
    --max_capacity_prompt $MAX_CAPACITY_PROMPT \
    --model_version LlaMA3_${METHOD}_${MAX_CAPACITY_PROMPT}_${TAG}
) 2>&1  | tee results_needle/logs/LlaMA3_${METHOD}_${MAX_CAPACITY_PROMPT}_${TAG}.log

Both LLaMA3 and Mistral2 inference support on single GPU.
model_provider: LLaMA3 or Mistral2
model_name: Path to your model. Support "Llama-3-8B-Instruct" "Mistral-7B-Instruct-v0.2" and for now.
step: The increase of context length.
method: Support PyramidKV, SnapKV, StreamingLLM, H2O.
max_capacity_prompt: Selected KV Size in each layer. （e.g. 128, 2048 in paper）. When method is "PyramidKV", given that the total number of KV remains unchanged, the specific KV length for each layer will be modified accordingly

To reproduce our results, run

bash scripts/scripts_needle/eval.sh

After inference, run

python scripts/scripts_needle/visualize.py

to draw the img, you should change FOLDER_PATH in visualize.py to your output path (the argument of --model_version in eval.sh).

Citation

If you find PyramidKV useful for your research and applications, please kindly cite using this BibTeX:

@misc{cai2024pyramidkv,
      title={PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling}, 
      author={Zefan Cai. and Yichi Zhang and Bofei Gao and Tianyu Liu and Keming Lu and Wayne Xiong and Yue Dong and Baobao Chang and Junjie Hu and Wen Xiao},
      year={2024},
      eprint={2406.02069},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Acknowledgement

Thanks [SnapKV] SnapKV: LLM Knows What You are Looking for Before Generation for providing open-source code to support the expansion of this project.

Zefan-Cai / PyramidKV

readme