Pick a path
ResonTech runs your model on bare GPU — you write the code, the platform handles provisioning, distributed execution, fault recovery, and artifact delivery. Two things you can do here:
| I want to… | Start here |
|---|---|
| Train a model (federated) | Submit via UI wizard |
| …or from Python | Submit via Python SDK |
| Deploy a model (inference) | Quick start |
| …or from Python | Submit via Python SDK |
| First time using the SDK? | Install the SDK |
| Want a runnable example? | vLLM examples gallery |
What's in these docs
Docs are organized in the order you'll need them:
- Get started — install the SDK, pick a path.
- Training — submit a federated training job (UI or SDK), watch it run, download results.
- Inference — deploy a hosted runtime, call
/predict, see a worked LLM example. - Reference — every config field, the file browser, custom classes, troubleshooting.
- Advanced & theory — how FL works on ResonTech, port a centralized recipe to FL, examples catalog, infrastructure.
- Product — what the platform does, who uses it, who we are.
If you're stuck on installation or your first submit, jump straight to Troubleshooting — it's organized by what you were trying to do.
What you change vs. what the platform handles
What you write
- Your model code (PyTorch, TensorFlow, JAX — any framework).
- For training: a one-round
fl_train_model()function. - For inference: a class with
__init__+predict(data: bytes). - A short YAML describing cluster + requirements.
What the platform handles
- GPU provisioning + worker selection.
- Distributed execution (NCCL / FSDP / Ray Serve, depending on the workload).
- Fault recovery from the last checkpoint.
- Artifact delivery back to your S3 bucket.
- Per-second billing — no idle cost when nothing's running.
The longer version (problem space, infrastructure, who this is for) lives on the How training & inference work page — read it when you're curious, not when you're trying to ship.