Awesome

AdaKV

Adaptive Budget Allocation across different attention heads based on their concentration degrees effectively improves budget utilization, thereby improving post-eviction generation quality.

Updates

• [2024.11.08 GQA Support] In response to numerous community requests, we’ve just uploaded a new testing branch, named test_gqa_support, which introduces support for Grouped Query Attention (GQA) under the methods "SnapKV," "PyramidKV," "Ada-SnapKV," and "Ada-PyramidKV." You can use the GQA_eval_longbench.sh script for direct evaluation on the LongBench benchmark. Detailed results will be released soon.
• [2024.11.15 GQA Results] We have integrated GQA support for Mistral-7B-Instruct-v0.2 in SnapKV, PyramidKV, and our Ada-KV. You can try it by running the GQA_eval_longbench.sh script! Preliminary results on a single A800 are provided in the table below. In future updates, we will include more comprehensive GQA test results in the appendix of our Ada-KV paper.
• [2024.12.12 Community Collaboration ] We’re collaborating with NVIDIA’s KVpress team to integrate Ada-KV into their excellent project. This effort also aims to build a foundation for future research on head-specific cache compression methods. See the draft PR for progress. Meanwhile, we’ve conducted broader evaluations of Ada-KV(Ada-SnapKV), including preliminary results on the 4K Ruler Benchmark (context compression with and without questions). For further insights into the evaluation, please refer to the kvpress repository.

Usage

Requirements

transformers==4.37.2
flash-attn==2.4.0

datasets
tiktoken
jieba
rouge_score

Installation

git clone https://github.com/FFY0/AdaKV
cd AdaKV
make i

Quick Start

# replace modeling with adakv
from adaptive_snapkv.monkeypatch.monkeypatch import replace_mistral_adaptive, replace_llama_adaptive
replace_mistral_adaptive()
replace_llama_adaptive()

model = AutoModelForCausalLM.from_pretrained(
    model_name_or_path,
    config=config,
    device_map=device_map,
    attn_implementation="flash_attention_2",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
)

# config hyperparameters
compress_args = {}
def config_compress(model, window_size=32, base_capacity=512, kernel_size=7, pooling="maxpool", floor_alpha=0.5, pyram_mode = False, beta = 20):
    model.model.config.window_size = window_size
    model.model.config.base_capacity = base_capacity
    model.model.config.kernel_size = kernel_size

    model.model.config.pooling = pooling
    model.model.config.floor_alpha = floor_alpha

    model.model.config.pyram_mode = pyram_mode
    model.model.config.pyram_beta = beta
    return model

model = config_compress(model, **compress_args)

Flattened Storage and Flash Attention Support

Considering varied cache length across heads, we implement a flattened storage layout of KV cache combined with flash_attn_varlen_func for efficent computation.

Regular MHA Cache Storage

Layer i:
    head0: (t00, t01, t02)
    head1: (t10, t11, t12)
    head2: (t20, t21, t22) 

past_key_value.update():

Layer i:
    head0: (t00, t01, t02, t03)
    head1: (t10, t11, t12, t13)
    head2: (t20, t21, t22, t23)

Note. tij means cache element of token j on head i in this case.

Flattened Cache Storage

The corresponding cuda code can be found in ./csrc/csrc/cuda_api.cu.

Layer i:
    (t00, t01, t02, t03) (t10, t11) (t20, t21, t22)

past_key_value.update():

Layer i:
    phase 0: malloc empty cache
    (_, _, _, _, _) (_, _, _) (_, _, _, _)

    phase 1: copy old value
    (t00, t01, t02, t03, _) (t10, t11, _) (t20, t21, t22, _)
    
    phase 2: insert new value
    (t00, t01, t02, t03, t04) (t10, t11, t12) (t20, t21, t22, t23)

Details about flash_attn_varlen_func can be found in Repo.

Peak Memory Footprint and Decoding Latency For Our Implementation:

Evaluations

LongBench without GQA Support

cd ./experiments/LongBench
bash runall.sh

Citation

If you found our work valuable, please cite:

@misc{feng2024adakvoptimizingkvcache,
      title={Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference}, 
      author={Yuan Feng and Junlin Lv and Yukun Cao and Xike Xie and S. Kevin Zhou},
      year={2024},
      eprint={2407.11550},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2407.11550}, 
}

Acknowledgement

We extend our gratitude to SnapKV and PyramidKV for their contributions of open-source code, which have significantly facilitated the advancement of this project.

Misc

Observation

Different attention heads within each layer of LLMs exhibit significant disparities in the degrees of attention concentration.

Therefore, we can improves budget utilization by dynamically allocating the budget across different attention heads within the same layer based on their concentration degrees.