Esacode

Downloads

The EsaSim worker

One program for every worker you run, here or remote: laptops, workstations and servers. Install it once as a service, then set the worker up from the dashboard. It connects out to EsaSim; nothing connects in.

EsaSim worker 0.13.2. On Windows one line installs everything, on a computer that has nothing installed. On Linux and macOS the first command downloads the source code and builds it, installing what the build needs; the second, with the code from Compute > Add a worker in your dashboard, pairs the worker and installs the service.
Windows0.13.2

An administrator PowerShell (Windows 10 or 11, 64-bit): Start, type PowerShell, Run as administrator

One line, with the code from Compute > Add a worker in your dashboard (the dashboard shows it ready to paste):

& ([scriptblock]::Create((irm https://esacode.com/install/worker.ps1))) -Code ABC123

Nothing needs to be on the computer first. The line downloads the worker program and checks it, gives it its own Python with PyTorch (with CUDA when the PC has an NVIDIA GPU) and the usual packages for machine learning, installs the Windows service "EsaSim worker", pairs it and starts it. Everything goes in C:\EsaSim\worker1 (add -Folder D:\other\path to choose). The first time takes a few minutes, mostly PyTorch. The worker updates itself from EsaSim.

The program alone (esasim-worker.exe, 4 MB, SHA-256 b297f3714c5564666212e002ce1342fbb7940f7e7f17d10df283f0ec1b6734ea)

Linux0.13.2

Terminal (systemd distributions: Ubuntu, Debian, Fedora, RHEL…), in the worker's folder

0. Make the worker's folder (one per worker; everything it makes stays there)

mkdir -p ~/esasim/worker1 && cd ~/esasim/worker1

1. Download and build, in that folder

curl -fsSL https://esacode.com/downloads/esasim-worker-src.tar.gz | tar -xz --strip-components=1 && scripts/setup.sh && scripts/build.sh

2. Pair (the dashboard shows it with your code)

sudo ./esasim-worker service install --code ABC123 --server https://esacode.com

3. Run: already running after step 2, and it starts with the worker. To stop it, and start it again:

sudo ./esasim-worker service stop
sudo ./esasim-worker service start
macOS0.13.2

Terminal (Apple silicon or Intel; presets without accelerators), in the worker's folder

0. Make the worker's folder (one per worker; everything it makes stays there)

mkdir -p ~/esasim/worker1 && cd ~/esasim/worker1

1. Download and build, in that folder

curl -fsSL https://esacode.com/downloads/esasim-worker-src.tar.gz | tar -xz --strip-components=1 && scripts/setup.sh && scripts/build.sh

2. Pair (the dashboard shows it with your code)

sudo ./esasim-worker service install --code ABC123 --server https://esacode.com

3. Run: already running after step 2, and it starts with the worker. To stop it, and start it again:

sudo ./esasim-worker service stop
sudo ./esasim-worker service start

Source code: esasim-worker-src.tar.gz (155 KB, SHA-256 6c71ca8e1e2058d0bba06424d2ac35f00e5ce069d163d28bdf38a41f08c6373a)

Paste both lines in the same window: the build leaves the program in the folder the first line moves into, so the second works right there. It asks for administrator rights itself: on Windows the usual “allow changes?” prompt, on Linux and macOS sudo. After pairing, everything else is set on the worker's page in the dashboard: tasks at the same time, which accelerators and presets it uses, working hours, pause and the data folder. Step-by-step guide, file locations and troubleshooting: Set up a worker.

Supported hardware

NVIDIAAMDIntelHuawei AscendTenstorrent

The worker finds accelerators by itself and reports them to the dashboard; each preset lists the hardware it runs on and the dashboard only sends a computer jobs it can run. Install the vendor's driver before the worker. No vendor SDK is needed to build it.

NVIDIAGeForce, RTX professional and data-center GPUs · found through the driver (NVML) · Windows and Linux
AMDRadeon, Radeon PRO and Instinct · found with rocm-smi (ROCm, Linux) and through Vulkan (Windows and Linux)
IntelArc, Data Center and integrated GPUs · found with xpu-smi and through Vulkan · Windows and Linux
AppleApple silicon GPUs · found through Metal · macOS; blocks use them with PyTorch (MPS) or MLX
Huawei AscendAtlas cards and servers · found with npu-smi · Linux
TenstorrentWormhole and Blackhole cards · found with tt-smi · Linux
CPU onlyPresets that need no accelerator run on any worker, macOS included

Workers with several GPUs: tasks share them, take one each, or take one of the groups you set (for example 1 + 3 + 4 GPUs of an 8-GPU worker), and each task sees only its own. esasim-worker scan lists what a card still needs, such as a driver. Workers with several GPUs

Shared GPU servers

For a team's multi-GPU Linux server split into compute units, with each job in its own Docker container, the server mode of the same worker is in development. You can already register servers and define compute units in the dashboard.

What you are installing

The worker only connects out over HTTPS with certificate checks always on. It runs only the presets built into it; the dashboard cannot send it commands to run. As a service it runs under a low-privilege account, and its access token is readable only by that account and administrators.

Coming next

  • Signed installers (MSI for Windows, .deb and .rpm for Linux, .pkg for macOS) wrapping the same service install
  • Self-update from the dashboard
  • Docker runner for shared GPU servers
  • Datasets stored where you choose, and jobs placed next to their data

Create your account first

Workers are paired with a code from the Computing page, so you need an account before installing.