Sign in Start now ›

Documentation

Docs

Train across many GPUs or serve a model behind one URL. Start with the overview for your workload.

Start here

Welcome

New here? See what ResonTech does, then pick your path: training or inference.

Install the SDK

pip install, two keys, one config. You’re connected in five minutes.

Inference overview

Paste a model id, get an endpoint. Both ways to serve, and what you pay for.

Training overview

Keep the training loop you have, and run it on as many GPUs as you want.

Inference

Serve a Hugging Face model through vLLM, or your own Python class, behind one URL and an API key.

Quick startFrom two files to a live endpoint you can call. Start here to try it now.Choosing hardwareWill your model fit a 16 GB card? Size the GPUs, the copies and the central server before you pay.Calling your modelYour endpoint is up. Send it chat requests, images or your own JSON, and handle what comes back.BenchmarksWe added a second GPU and measured: twice the answers, and the first word up to 12× sooner.

Serve vLLM models

How it worksTwo files turn a Hugging Face LLM into an OpenAI-compatible API. See what each one does.serve_module.pyThe one Python file where you name your model and decide how it runs.inference.yamlTell the platform what hardware your model needs, and keep it in step with your Python.LLMConfig optionsModel won’t boot, or the context is too short? These are the settings that fix it.

vLLM examples

Examples galleryChat, coding, vision, RAG and speech bundles. Submit one as it is, change it later.Qwen3-8B guarded RAG chatA full RAG stack with a safety guard, all on one 24 GB card.Qwen2.5-VL-32B visionTurn hard scans, handwriting and messy tables into clean JSON.

Serve custom models

How it worksAny model that runs in Python can be an endpoint. You write one class and skip the web server.serve.pyThe class you write: two required methods and three optional ones.inference.yamlPoint the platform at your class, pass it arguments and ship your weights.

Custom model examples

Examples gallerySegmentation, image generation, text-to-speech and classifiers, ready to submit.SAM 2.1 segmentationClick a point on an image and get the object’s mask back, on a 12 GB card.

Inference SDK

Submit via SDKDeploy from a script or CI in one call and get the endpoint and key back in Python.

Training

Each GPU trains on its own shard of your dataset; the central server merges the weights after every round.

Submit via UI wizardLaunch a training job from the browser in seven steps, with every setting explained.Choosing hardwareWill an 8B fine-tune fit your cards? Pick the GPUs, how many, and the server that merges them.Dataset format & shardingSplit your data so every GPU learns well. Layouts and checks for images, text and audio.Track jobs & download resultsFollow each round as it runs, then download the trained model.FL examples galleryTwelve working training jobs to copy, from vision and language to speech and medical imaging.BenchmarksThree A40s against one RTX 5090 on the same job: a third of the time, about half the cost.

Training SDK

Submit via SDKStart a training job from Python in one call and watch its rounds from your notebook.Training hyperparametersRounds, epochs, learning rate and the rest, and how to check the config before you launch.Custom training classesBring your own model, training step or checkpoint logic when the defaults don’t fit.

Advanced & theory

FL internals, porting an existing recipe, and the central servers that coordinate your jobs.

How FL works on ResonTechWhat happens inside each round, and the tuning lessons from our own runs.Port a centralized recipe to FLAlready have a train.py? Move it onto many GPUs with 50–100 lines of changes.Central serversThe machine that coordinates your GPUs: what it does, and how to get one of your own.

Using an AI assistant? Point it at /llms.txt, a plain-text index of every page.