DSC: A Tensor Library & Inference Framework for Machine Learning

DSC is a powerful, PyTorch-compatible tensor library designed specifically for machine learning models. It offers a low-level API that is C-compatible, wrapped in a modern Python interface that resembles popular libraries like NumPy and PyTorch, while also providing notable usability enhancements.

Key Features

• Intuitive API: The DSC Python API is designed to be user-friendly, making it easy for developers familiar with NumPy and PyTorch to get started.

• Built-In Neural Network Support: DSC includes nn.Module built-in functionality, making the transition from PyTorch to DSC straightforward. For illustrative examples, refer to the examples directory.

• Support for Multiple Backends: Whether utilizing CPU or CUDA, DSC accommodates various backends effortlessly. Users can switch between backends by using the dsc.set_default_device('...') command, ensuring minimal code changes.

• Minimal External Dependencies: Efficiency is a key focus of DSC. It requires no external libraries, and core operations are crafted from the ground up in portable C++, ensuring high portability of the code.

• Custom Memory Allocation: DSC features a custom memory allocator that pre-allocates memory only once, eliminating the overhead of frequent calls to malloc() and free(). Additionally, users can opt for a linear allocator for further optimization.

Quick Start Guide

Getting started with DSC is straightforward. Below are the primary requirements:

• A C++20-compatible compiler
• GNU Make for building

To set up on a Linux-based system, use the following commands:

sudo apt update
sudo apt install build-essential

For installation from source:

git clone git@github.com:nirw4nna/dsc.git
cd dsc/
python3 -m venv venv
source venv/bin/activate
python3 -m pip install -e .

To compile the C++ library:

make clean; make shared DSC_FAST=1

Verify the installation by performing a simple operation:

python3 -c "import dsc; x = dsc.arange(10); print(x)"

Utilizing the CUDA Backend

For GPU acceleration on NVIDIA GPUs, ensure that the CUDA Toolkit is installed. To build the C++ library with CUDA support:

make clean; make shared DSC_FAST=1 DSC_CUDA=1

Run a simple GPU operation to check if the CUDA backend is functioning:

python3 -c "import dsc; x = dsc.arange(10, device='cuda'); print(x)"

Testing Framework

DSC employs pytest for unit testing against NumPy, the reference for correctness. To run all tests:

cd python/tests/
pytest -s test_ops.py --no-header --no-summary -q

Compile DSC with DSC_FAST=1 for optimized testing.

DSC stands out as a compact yet efficient framework for machine learning inference, ensuring both performance and ease of use.