• Home
• GitHub
• License (BSD)
• Getting Started
• Install
• Quickstart
• Tutorials
• More Examples
• API Documentation
• API Reference
• Extensions

Installing

dfx runs on dotnet, a cross-platform, open-source platform supported on Linux, macOS, and Windows.

There are various ways in which you can run dfx, the main ones being: interactive notebooks supporting Visual Studio Code and Jupyter; running in a REPL; running script files; and compiling, packing, and publishing performant binaries.

Interactive Notebooks and Scripts

You can use dfx in dotnet interactive notebooks in Visual Studio Code or Jupyter, or in F# scripts (.fsx files), by referencing the package as follows:

// Use one of the following three lines
#r "nuget: dfx-cpu" // Use the latest version
#r "nuget: dfx-cpu, *-*" // Use the latest pre-release version
#r "nuget: dfx-cpu, 1.0.10" // Use a specific version
open Dfx

Dotnet Applications

You can add dfx to your dotnet application using the dotnet command-line interface (CLI).

For example, the following creates a new F# console application and adds the latest pre-release version of the dfx-cpu package as a dependency.

dotnet new console -lang "F#" -o src/app
cd src/app
dotnet add package --prerelease dfx-cpu
dotnet run

Packages

We provide several package bundles for a variety of use cases.

• dfx-cpu
  Includes LibTorch CPU binaries for Linux, macOS, and Windows.
• dfx-cuda-linux / dfx-cuda-windows
  Include LibTorch CPU and CUDA GPU binaries for Linux and Windows. Large download.
• dfx-lite
  Includes the Torch backend but not the LibTorch binaries.

Using local LibTorch binaries (optional)

You can combine the dfx-lite package bundle with existing local native binaries of LibTorch for your OS (Linux, Mac, or Windows) installed through other means.

LibTorch is the main tensor computation core implemented in C++/CUDA and it is used by PyTorch in Python and by other projects in various programming languages. The following are two common ways of having LibTorch in your system.

• If you use Python and have PyTorch installed, this comes with LibTorch as a part of the PyTorch distribution. If your GPU works in this PyTorch installation without any issues, it will also work in dfx.
• You can download the native LibTorch package without Python by following the get started instructions in the PyTorch website, and extracting the downloaded archive to a folder in your system.

Before using the Torch backend in dfx, you will have to add an explicit load of the LibTorch native library, which you can do as follows. In order to find the location of LibTorch binaries, searching for libtorch.so in your system might be helpful. Note that this file is called libtorch.so in Linux, libtorch.dylib in macOS, and torch.dll in Windows.

open System.Runtime.InteropServices
NativeLibrary.Load("/home/user/anaconda3/lib/python3.8/site-packages/torch/lib/libtorch.so")

Backends and Devices

dfx currently provides two computation backends.

• The Torch backend is the default and recommended backend based on LibTorch, using the same C++ and CUDA implementations for tensor computations that power PyTorch. On top of these raw tensors (LibTorch's ATen, excluding autograd), dfx implements its own computation graph and differentiation capabilities. This backend requires platform-specific binaries of LibTorch, which we provide and test on Linux, macOS, and Windows.

• The Reference backend is implemented purely in F# and can run on any hardware platform where dotnet can run (for example iOS, Android, Raspberry Pi). This backend has reasonable performance for use cases dominated by scalar and small tensor operations, and is not recommended for use cases involving large tensor operations (such as machine learning). This backend is always available.

Configuration of Default Backend, Device, and Tensor Type

Selection of the default backend, device, and tensor type is done using dfx.config.

• Dtype choices available: BFloat16, Bool, Byte, Float16, Float32, Float64, Int16, Int32, Int64, Int8

• Device choices available: CPU, GPU

• Backend choices available: Reference, Torch

For example, the following selects the Torch backend with single precision tensors as the default tensor type and GPU (CUDA) execution.

open Dfx

dfx.config(dtype=Dtype.Float32, device=Device.GPU, backend=Backend.Torch)

The following selects the Reference backend.

dfx.config(backend=Backend.Reference)

A tensor's backend and device can be inspected as follows.

let t = dfx.tensor [ 0 .. 10 ]

let device = t.device
let backend = t.backend

Tensors can be moved between devices (for example from CPU to GPU) using Tensor.move. For example:

let t2 = t.move(Device.GPU)

Developing dfx Libraries

To develop libraries built on dfx, you can use the following guideline to reference the various packages.

• Reference the dfx-Dfx.Core and dfx-Dfx.Data packages in your library code.
• Reference the dfx-Dfx.Backends.Reference package in your correctness testing code.
• Reference the dfx-Dfx.Backends.Torch and libtorch-cpu packages in your CPU testing code.
• Reference the dfx-Dfx.Backends.Torch and libtorch-cuda-linux or libtorch-cuda-windows packages in your (optional) GPU testing code.

© Copyright 2021, DiffSharp and dfx Contributors.