Skip to content

Latest commit

 

History

History
102 lines (81 loc) · 5 KB

README.md

File metadata and controls

102 lines (81 loc) · 5 KB
Raw

LLM-PQ

Official Repo for LLM-PQ: Serving LLM on Heterogeneous Clusters with Phase-Aware Partition and Adaptive Quantization

| Poster | Full Paper｜

  • LLM-PQ argues that the assumption of infinite requests in LLM serving is not necessarily valid.
  • LLM-PQ emphasizes the importance of efficiently processing workload-similar predetermined offline batch processing tasks
  • But also maximizing the utilization of GPUs acquired at different points in time (Heterogenous GPU Serving).
  • Utilize the overall resources towards behaviors of quantization for the same workload in different phases (Ada-Q).
  • Especially, LLM-PQ is a workload-centric and device-agnostic serving framework, taking both workload information and device information for strategy.

In this version, we don't have a chatbot, but a flexgen-like one-time running script.

Before You Proceed

  • Due to historical reasons (this repository was initially built between March and June 2023), LLM-PQ's pipeline is built on top of PipeEdge. As a result, its performance may be limited compared to the latest pipeline implementations, such as TGI. However, this also ensures a fair comparison with PipeEdge.

Install

LLM-PQ is implemented in a top-down view, where

  • LLM-PQ: Provides the distributed runtime and optimizer for the better serving plan
  • QLLM: the customized LLM workload and its quantized version
  • LPTorch: the innermost quantization support for the LM, implement different quantization schemes.

Due to a similar reason, latter two's performance is not a SOTA. If this repo / paper is getting popular 🤑🤑🤑, we will consider merging / updates the later two.

Docker (Recommended)

You can use the docker file under the dockerfiles. We also provide the pre-built image with data insides:

    docker pull springtonyzhao/llmpq:v100 # v100 (the one who required from scratch build of bitsandbytes)
    docker pull springtonyzhao/llmpq:a100 # A100

Manual

    git clone --recursive https://github.com/tonyzhao-jt/llm_pq.git
    python3 pip install -e .

Careful: use GPU with cap <= 70 require recompile of bitsandbytes. We done it for u in setup.py, but if not, please run the update.sh in the 3rd_party of LPTorch to manually compile and install the bitsandbytes.

Possible errors

  • BuilderConfig 'allenai--c4' not found. Available: : please change the data load script in GPTQ to
traindata = load_dataset(
    'allenai/c4', data_files={'train': 'en/c4-train.00000-of-01024.json.gz'}, split='train'
)
  • ERROR: Could not install packages due to an OSError:: when pip install -e ., you can just install it again and the problem can be solved.

Optimizer Setup

llm_pq's optimizer utilizes the support from the gurobi. To use gurobi, put the web license under /opt/gurobi/ or under configs:

Else you will get:

    ERROR:llm_pq: Please install gurobi and put the license file under /opt/gurobi/

Reproduce Results in Paper

Scripts

We provide all the scripts used in the paper under the scripts/ folder. Currently, we evaluate the performance and accuracy separately by performing the following:

  • Evaluating mixed-precision using modified scripts from GPTQ.
  • Evaluating performance using the distributed runtime.

Please note that in the current version, we load the model layer by layer and do not require any additional storage for weight saving. However, this loading process might be relatively slow.

Graphing

We provide all the graphing scripts under the notebook/ folder.

  • For cost model, you need to profile gtruth for prediction error est.

TODOs if 🌟

  1. Faster Loading: We are going to add scripts to the distributed runtime and quantization part to make it can be fast deployed in runtime.
  2. Better Pipeline: Replace PipeEdge's pipeline with sth better.
  3. More efficient model structure: The existing model structure is directly adopted from the old transformer lib, introducing many unnecessary ops that could be reduced. But also, we only provide BLOOM / OPT for the moment, which could be also improved.
  4. Deployment: Wrap it with a chatbot.

Citation

If you use LLM-PQ for your research, please cite our paper:

@inproceedings{10.1145/3627535.3638480,
author = {Zhao, Juntao and Wan, Borui and Wu, Chuan and Peng, Yanghua and Lin, Haibin},
title = {POSTER: LLM-PQ:Serving LLM on Heterogeneous Clusters with Phase-Aware Partition and Adaptive Quantization},
year = {2024},
isbn = {9798400704352},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3627535.3638480},
doi = {10.1145/3627535.3638480},
pages = {460–462},
keywords = {LM serving, heterogenous cluster, quantization},
series = {PPoPP '24}
}