The guides on this site are written for one situation: you are paying for a GPU by the hour, so every hour spent on setup, lost files or an instance nobody stopped is money gone. Each guide covers one task on DeployGPU from start to finish, with the commands to copy and the platform details that generic tutorials leave out, such as which username each provider uses and what stopping an instance deletes.
Where to start
New to renting GPUs? Read how to rent a cloud GPU first. It walks the whole loop in 9 steps: picking a GPU by memory, adding credits, launching, connecting and stopping.
Connecting to the machine is the next thing that goes wrong, usually because of the username or the key. SSH into a cloud GPU covers keys, the username per provider, ~/.ssh/config, VS Code and tunnels.
Before your first stop, read save your work, then stop your GPU. Stopping deletes the instance and its disk, and that guide is about not learning it the hard way.
Your first session on the machine
Once connected, check a new GPU in the first five minutes with nvidia-smi before you copy anything onto it. Keeping jobs running after you disconnect covers tmux, nohup and systemd, and SSH port forwarding reaches Jupyter, TensorBoard or a model server without opening a port. Getting software and data onto the machine has 4 guides: installing PyTorch with CUDA, running Docker with GPUs, downloading Hugging Face models fast and moving data to and from a cloud GPU. VS Code Remote-SSH turns the VM into your editor.
Running models
For open-weight LLMs, self-hosting an LLM on a rented GPU is the overview: which model, which GPU, which serving stack. From there, deploy vLLM on a cloud GPU sets up an OpenAI-compatible API on a GPU VM, and two comparisons help you choose: vLLM, Ollama, llama.cpp or SGLang, and Ollama Cloud or your own GPU. The engine guides cover llama.cpp server for GGUF files and SGLang, and the quantization guide explains which formats each one runs. For other workloads, see Whisper transcription and models and custom nodes in ComfyUI. The models section has the memory math per model.
Training and fine-tuning
Fine-tuning on cloud GPUs is the starting point, with step-by-step guides for Unsloth and image-model LoRAs. Multi-GPU training covers torchrun, DDP and FSDP on one 8-GPU VM, and checkpoint and resume keeps a run from dying with the instance.
Automating with the API
If you launch GPUs from code, handling provisioning failures covers failure codes, the 503 cooldown and why a create call must never be retried blindly. Batch GPU jobs from a script runs a job end to end, and API cost guardrails stops machines that would otherwise keep billing. The API page has a complete launch script, and GPU cloud aggregators explained covers how a multi-provider platform like this one works.
Every guide
When something breaks rather than needs setting up, the troubleshooting section is organised by the error message you see.