gQuant - GPU Accelerated Graph Computation for Quantitative Analyst Tasks

NOTE: For the latest stable README.md ensure you are on the master branch.

What is gQuant?

gQuant is a collection of open-source GPU accelerated Python tools and examples for quantitative analyst tasks, built on top of the RAPIDS AI project, Numba, and Dask.

The examples range from simple accelerated calculation of technical trading indicators through defining workflows for interactively developing trading strategies and automating many typical tasks.

The extensibility of the system is highlighted by examples showing how to create a dataframe flow graph, which allows for easy re-use and composability of higher level workflows.

The examples also show how to easily convert a single-threaded solution into a Dask distributed one.

These examples can be used as-is or, as they are open source, can be extended to suit your environments.

gQuant jupyterlab extension

The gQuant Juyterlab extension provides the user interface to build the dataframe flow TaskGraph easily. It takes advantage of the open sources projects like jupyterlab, ipywidget, React and D3. It features:

1. Takes full advantage of the JupyterLab project that the extension adds commands to Jupyterlab context menu, command palette and bind them with keyboard shortcuts to speed up the productivity.
2. Define a new TaskGraph file format .gq.yaml that can be edited in the Jupyterlab.
3. Visually presents the TaskGraph as a DAG graph. Users can zoom in and out, freely move the nodes around, and make connections between nodes.
4. Use the special Ouput Collector to gather the results and organize them in a tab widget. The IPython rich display is fully supported.
5. Visually shows the progress of graph evaluation and computation dependence.
6. Automatically generate the UI elements to edit and validate the Node configuration given the configuration JSON schema. It exposes the function API in a user-friendly way. User can change the configuration and re-run the computation to test out the hyperparameters easily.
7. Dynamically compute the input-output ports compatibility, dataframe columns names and types, ports types to prevent connection errors.
8. Nodes can have multiple output ports that can be used to generate different output types. E.g. some data loader Node provides both cudf and dask_cudf output ports. The multiple GPUs distributed computation computation is automatically enabled by switching to the dask_cudf output port.
9. Provides the standard API to extend your computation Nodes.
10. The composite node can encapsulate the TaskGraph into a single node for easy reuse. The composite node can be exported as a regular gQuant node without any coding.

Prerequisites

• NVIDIA Pascal™ GPU architecture or better.
• CUDA 9.2 with driver v396.37+ or CUDA 10.0 with driver v410.48+.
• Ubuntu 16.04 or 18.04.
• NVIDIA-docker v2+.

Download data files

Run the following command at the project root diretory

bash download_data.sh

Install

gQuant source code can be downloaded from GitHub.

• Git clone source code:

$ git clone https://github.com/rapidsai/gQuant.git

• Build and run the container:

$ cd gQuant/docker && . build.sh

When building the container, you can run gQuant in two modes: dev or prod. In the dev mode, please check the README file in gquantlab directory to install the plugins and Python libraries.

In the production mode, you can launch the container by following command and start to use it

$ docker run --runtime=nvidia --rm -it -p 8888:8888 -p 8787:8787 -p 8786:8786 gquant/gquant:[tag from the build]

Example notebooks

Example notebooks, tutorial showcasing, can be found in notebooks folder.

System environment

There are a few system environment that the user can overwrite.

The custom module files are specified in the gquantrc file. GQUANT_CONFIG enviroment variable points to the location of this file. By default, it points to $CWD\gquantrc.

In the example gquantrc, system environment variable MODULEPATH is used to point to the paths of the module files. To start the jupyterlab, please make sure MODULEPATH is set properly.

For example, if you want to start the jupyterlab in the gQuant root directory.

MODULEPATH=$PWD/modules jupyter-lab --allow-root --ip=0.0.0.0 --no-browser --NotebookApp.token=''

Or, if you want to start the jupyterlab in the gquantlab directory.

GQUANT_CONFIG=../gquantrc MODULEPATH=$PWD/../modules jupyter-lab --allow-root --ip=0.0.0.0 --no-browser --NotebookApp.token=''