KV cache compression for high-throughput LLM inference

Paper | Blog | X/Twitter

---

This is a (messy) fork of vLLM v0.6.0 showcasing our new KV cache compression method that increases throughput for memory-constrained LLM deployments.

Setup

It is recommended to run within the NVIDIA PyTorch image:

docker run --gpus all -it --rm --ipc=host nvcr.io/nvidia/pytorch:24.04-py3

Install from source:

cd vllm-kvcompress/
pip install -e .

Alternatively, the prebuilt wheel can be used for x86 architectures:

pip install https://pub-ff08b7559526447fb14dd52ec4fac7c7.r2.dev/17da8eb/build/sm_89/vllm-0.6.0%2Bcu124-cp310-cp310-linux_x86_64.whl

Inference

The inference server can be launched with:

export model=meta-llama/Meta-Llama-3.1-8B-Instruct
vllm serve $model --enforce-eager --enable-kvc

Requests can then be sent with

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
    "prompt": "San Francisco is a",
    "max_tokens": 7,
    "temperature": 0,
    "max_cache_tokens": 128,
    "protected_window_size": 32,
    "compress_once": false
  }'

Compression can be configured per-request by setting the following additional sampling parameters:

• max_cache_tokens - Configure the max number of KVs to retain in cache for this sequence, computed as num_layers * num_kv_heads * max_cache_tokens
• protected_window_size - The window of final tokens for this sequence whose KVs will be protected during compression.
• compress_once - If set, the sequence will only be compressed during the first compression iteration after its prefill.

Running Experiments

LongBench

cd vllm-kvcompress/experiments/

To run experiments with a limited observation window (KVC-w):

export dataset=narrativeqa model=llama3 w=8 cache_size=128
python run_longbench.py \
  --dataset $dataset \
  --model $model \
  --protected-window-size $w \
  --prefill-metric-collection-window-size $w \
  --max-cache-tokens $cache_size

To run experiments with full query-range aggregation (KVC-full):

python run_longbench.py \
  --dataset $dataset \
  --model $model \
  --protected-window-size 32 \
  --metric-collection-buffer-size 10 \
  --prefill-metric-collection-window-size 33000 \
  --prefill-metric-collection-block-size 1024 \
  --no-maxpool-metrics \
  --gpu-mem-util 0.6 \
  --max-cache-tokens $cache_size

Note: Aggregating over the full query-range requires significant memory and should be run on an H100 or comparable GPU to avoid OOMs. Lowering gpu-mem-util will save more GPU memory for the aggregation and lowering prefill-metric-collection-block-size will lower the required memory for the aggregation, at the expense of longer execution time.

Experiments can be run with continual compression (compressing during decoding as well as on prefill) by adding the --continual-compression flag. To reproduce results in the paper, --compression-rate can be used to limit cache size instead of --max-cache-tokens:

export cr=64
python run_longbench.py \
  --dataset $dataset \
  --model $model \
  --protected-window-size $w \
  --prefill-metric-collection-window-size $w \
  --continual-compression \
  --compression-rate $cr

Run scripts used for our experiments can be found in experiments/scripts.

Benchmark Throughput

cd vllm-kvcompress/

Run vLLM's benchmarking script with:

export model=meta-llama/Meta-Llama-3.1-8B-Instruct \
  max_model_len=19000 input_len=6000 cr=64
python3 benchmarks/benchmark_throughput.py \
  --model $model \
  --max-model-len $max_model_len \
  --enforce-eager \
  --num-prompts 256 \
  --input-len $input_len \
  --output-len 500 \
  --protected-window-size 32 \
  --compression-rate $cr \
  --enable-kvc

Run scripts used for our experiments can be found in benchmarks/scripts.

Citation

If you use this work in research/projects of your own, please cite our paper:

@misc{rehg2024kvcompresspagedkvcachecompression,
      title={KV-Compress: Paged KV-Cache Compression with Variable Compression Rates per Attention Head},
      author={Isaac Rehg},
      year={2024},
      eprint={2410.00161},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.00161},
}