SDK
The IOTA SDK is coming soon
The IOTA SDK will let you configure a workload and run it across disaggregated compute directly, with the coordination, execution, and recovery handled for you.
The SDK principle
You define the workload. IOTA runs it.
You should not have to become a distributed-systems team to train across disaggregated compute. The SDK keeps you working at the level of your workload, and takes on the coordination underneath.
Developer workflow
Define your model, data, and training objective the way you already do. Connect to IOTA, and the same workload runs across disaggregated, variable compute, without you rewriting it for the system underneath.
From code to training
From a few lines of code to a running system.
You describe the workload in code, IOTA takes it from there. It configures, distributes, and runs the workload, and reports back as it goes. Our own groundbreaking results reproduce in around 500 lines.
# Illustrative only. Replace with real IOTA SDK syntax.
workload = iota.Workload(
model=my_model,
dataset=my_dataset,
objective="...",
)
training = iota.Training(
workload=workload,
config=my_config,
)
# IOTA handles distribution, execution, and recovery
run = training.start()
run.report()Pseudo-code placeholder onlyBuilt for technical control
Distributed underneath. Yours to control on top.
The SDK handles the hard parts of distribution without taking the controls off you. You decide how the workload is configured and run, IOTA makes it work across the compute beneath it.
01
Workload configuration
Define the model, data, objective, and training setup, and change them as you iterate. The configuration is yours to own and version, so a run is reproducible and a change is a deliberate edit, not a black box you have to trust.
What IOTA handles underneath
The systems work is ours, not yours.
Running training across disaggregated, variable compute is a hard systems problem. The SDK takes it on, so your workload holds together and stays correct without you managing the machinery.
01
Orchestration
We decide what runs where, and keep many moving parts agreeing on the state of the work. As machines join, leave, or fall behind, the plan updates around them, so every part of the run knows what it is doing and the whole stays in step without you directing it.
Designed for your workflow
Built to fit how you already build.
The SDK slots into existing setups instead of replacing them, whether you train models, run a platform, or build research tools.
Existing training workflows
Bring your current training setup and run it across disaggregated compute, without rewriting it for a new system.
Platform integration
Build IOTA into a platform or product, so training becomes something you can offer your own users.
Research and ML tooling
Integrate IOTA into research and experimentation tooling, and run demanding workloads on flexible capacity.
Get started
Start building on IOTA.
Read the docs to see how it works, or talk to the team about fitting it to your setup.