# ResonTech Documentation > ResonTech is a GPU compute platform: deploy inference runtimes (OpenAI-compatible > vLLM serving, or a custom Python /predict contract) and run federated training > jobs across a distributed GPU fleet. You ship a small bundle (inference.yaml + > scripts/), the platform provisions workers and returns an endpoint URL plus an > X-API-Key predict key. These docs cover the Python SDK, both serving contracts, > sizing/hardware guidance, and ready-to-run examples. Every page below is prerendered static HTML, fetchable without JavaScript. Runnable example jobs live at https://github.com/resontech-dev/examples (inference_jobs/vllm, inference_jobs/predict, training/nvflare). ## Get started - [Welcome](https://beta.reson.tech/docs/welcome): What ResonTech is, how it works, and where to start: training or inference. - [Install the SDK](https://beta.reson.tech/docs/install-sdk): pip install, two keys from your profile, one config. Five minutes to a working session. ## Inference - [Overview](https://beta.reson.tech/docs/inference-intro): A Hugging Face model through vLLM, or any Python you write, served behind one URL and an API key. - [Quick start](https://beta.reson.tech/docs/inference-quickstart): Submit a job, get a URL and an API key, call your model. Both runtime types on one page. - [Choosing hardware](https://beta.reson.tech/docs/inference-hardware): GPU memory, replicas, context and users: size the workers and the central server before you submit. - [Calling your model](https://beta.reson.tech/docs/inference-predict): One endpoint, one key: OpenAI calls for vLLM runtimes, bytes or JSON for custom ones, plus limits, errors and callbacks. - [Benchmarks](https://beta.reson.tech/docs/inference-benchmarks): One vs two RTX 5070 Ti GPUs on the same model: 2× the throughput, up to 12× faster first word. ## Inference — Serve vLLM models - [How it works](https://beta.reson.tech/docs/vllm-how-it-works): One serve_module.py and one inference.yaml become an OpenAI-compatible endpoint. - [serve_module.py](https://beta.reson.tech/docs/vllm-serve-module): The Python file that declares your model: one LLMConfig per engine, exported as app. - [inference.yaml](https://beta.reson.tech/docs/vllm-inference-yaml): The vLLM-mode yaml: required fields, the vllm: block, and the cluster knobs that matter. - [LLMConfig options](https://beta.reson.tech/docs/vllm-llmconfig): The knobs that decide whether your model boots and how much context it serves, VRAM math included. ## Inference — vLLM examples - [Examples gallery](https://beta.reson.tech/docs/vllm-examples): Ready-to-run vLLM jobs for chat, coding, vision, RAG and speech. Each is a folder you submit as it is. - [Qwen3-8B guarded RAG chat](https://beta.reson.tech/docs/vllm-example-qwen3-rag): Four engines on one 24 GB GPU: chat, moderation guard, embeddings and a reranker behind one endpoint. - [Qwen2.5-VL-32B vision](https://beta.reson.tech/docs/vllm-example-qwen-vl): Document vision for hard scans, handwriting and dense tables, with structured JSON out. ## Inference — Serve custom models - [How it works](https://beta.reson.tech/docs/predict-how-it-works): Ship one Python class, get a /predict endpoint. The platform owns HTTP, auth and scaling. - [serve.py](https://beta.reson.tech/docs/predict-serve-py): The predictor contract: __init__ and predict(data: bytes) → dict, plus three optional hooks. - [inference.yaml](https://beta.reson.tech/docs/predict-inference-yaml): The custom-mode yaml: module, class_name, init_args, weights and cluster sizing. ## Inference — Custom model examples - [Examples gallery](https://beta.reson.tech/docs/predict-examples): Ready-to-run predictor jobs: segmentation, image generation, TTS and classification. - [SAM 2.1 segmentation](https://beta.reson.tech/docs/predict-example-sam2): Point-prompted image segmentation on any 12 GB card: click a point, get masks. ## Inference — Inference SDK - [Submit via SDK](https://beta.reson.tech/docs/inference-sdk): rt_submit_inference() in detail: generator mode, your own bundle, weights and sample data. ## Training - [Overview](https://beta.reson.tech/docs/training-intro): Keep your training loop, wrap it in one function, and it runs across as many GPUs as you want. - [Submit via UI wizard](https://beta.reson.tech/docs/training-ui): The 7-step wizard that launches a training job from the browser. - [Choosing hardware](https://beta.reson.tech/docs/training-hardware): VRAM per client by method, how many clients, and what the FL server needs before you submit. - [Dataset format & sharding](https://beta.reson.tech/docs/dataset-format): One zip per worker: how to split, the thresholds a shard must meet, layouts by data type, and a pre-upload checklist. - [Track jobs & download results](https://beta.reson.tech/docs/training-monitor): Watch the rounds, read the logs, and download the final model. - [FL examples gallery](https://beta.reson.tech/docs/examples-fl): Twelve FL recipes across vision, language, speech, embeddings, medical and image generation. - [Benchmarks](https://beta.reson.tech/docs/training-benchmarks): 3× A40 against 1× RTX 5090 on one training run: 3× faster and 47% cheaper, same result. ## Training — Training SDK - [Submit via SDK](https://beta.reson.tech/docs/training-sdk): One rt_submit() call: define a model, ship the shards, watch it run. - [Training hyperparameters](https://beta.reson.tech/docs/training-config): TrainingConfig fields, the extra bag, federation knobs, and how to preview the rendered configs. - [Custom training classes](https://beta.reson.tech/docs/custom-classes): Override the model, executor or persistor, and preview the generated scripts. ## Advanced & theory - [How FL works on ResonTech](https://beta.reson.tech/docs/how-fl-works): FedAvg plumbing, the per-round loop, the persistor lifecycle and tuning lessons. - [Port a centralized recipe to FL](https://beta.reson.tech/docs/port-to-fl): Already have train.py? Wrap your model in fl_train_model() and ship it as an FL job. - [Central servers](https://beta.reson.tech/docs/central-servers): What the server that coordinates your GPUs does, how it is chosen, and how to reserve one. ## Optional - [Examples repository](https://github.com/resontech-dev/examples): complete submittable jobs: vLLM chat, coding, vision and RAG bundles, custom predictors (segmentation, TTS, image generation), and twelve federated-training templates. - [Dashboard](https://beta.reson.tech/dashboard): submit and manage runtimes (sign-in required).