TrainPods

TrainPods are Podstack’s raw, on-demand NVIDIA GPU compute: rent a GPU pod, SSH in with your own key, move data in and out, and run training or fine-tuning — provisioned, connected, and controlled entirely from the podstack CLI.

If you want a machine with GPUs that behaves like a plain Linux box you SSH into, TrainPods are it. Pick a GPU, launch it, connect, and start working. You pay by the hour only while the instance runs.

What you can do

• Provision GPUs on demand — one command launches an H100, A100, L40S, or any other GPU in the catalog, on spot or on_demand pricing.
• SSH in with your local key — Podstack registers only your public key; the private key never leaves your machine. podstack gpu instances ssh <id> drops you into a shell.
• Move data in and out — podstack gpu instances cp for SCP transfers, or podstack send / podstack receive for large, resumable, relay-based transfers.
• Expose services locally — tunnel TensorBoard, a training dashboard, or an inference server to localhost over SSH with podstack gpu instances expose.
• Run managed fine-tuning — hand Podstack a dataset and a base model with podstack train create and let the platform run a LoRA/QLoRA job for you — no instance management required.

What’s available today

TrainPods are driven through the podstack CLI, and every step below works right now:

• Provision GPUs on demand — podstack gpu instances create (or the interactive podstack gpu launch) starts an instance; browse the catalog and rates first with podstack gpu types list and podstack gpu pricing, on spot or on_demand pricing.
• Manage instances — podstack gpu instances list, get, and delete cover the full lifecycle.
• SSH with your own key — register keys with podstack gpu keys create / list / delete, then podstack gpu instances ssh <id> drops into a shell. Only your public key is registered; the private key stays local.
• Move data in and out — podstack gpu instances cp for SCP transfers, or podstack send / podstack receive for large, resumable, relay-based transfers.
• Expose services locally — podstack gpu instances expose <id> <port> tunnels TensorBoard, a dashboard, or an inference server to localhost.
• Run managed fine-tuning — podstack train create runs LoRA/QLoRA jobs with no box to manage; podstack train list / get / cancel / events track them and podstack files upload sends the dataset. See Fine-tuning.
• Per-hour, wallet-based billing — you pay by the hour only while an instance runs. See Pricing & billing.

Not yet generally available: an in-portal GPU marketplace (browse and order instances from the dashboard) is built but currently gated behind a feature flag, so it may not appear in your portal. Today, provision TrainPods through the CLI.

Use cases

• Custom training on rented hardware — an ML engineer runs podstack gpu launch, picks an H100, SSHes in, and runs their own multi-GPU training script with full root control.
• Fine-tune an LLM, hands-off — a data scientist uploads a JSONL dataset, runs a managed LoRA job with podstack train create --budget 25, and collects the fine-tuned model id — no instance to manage.
• Bring your own environment — a PhD student launches a raw GPU box, installs a specific CUDA and framework stack, and runs experiments that don’t fit a prebuilt template.
• Cheap overnight batch run — a cost-conscious team launches a spot-priced instance for an overnight job and deletes it in the morning to stop billing.
• Watch training remotely — an engineer starts TensorBoard on the instance and runs podstack gpu instances expose <id> 6006 to view the curves at localhost over SSH.
• Move a large dataset to the GPU — a researcher pushes a multi-gigabyte dataset with podstack send and pulls checkpoints back with podstack receive, resuming if the link drops.

TrainPods vs QuickPods

Podstack offers two ways to get GPUs. Pick based on how much of the environment you want to manage.

New to container pods and templates instead? See Compute → Pods.

Get started

Work through these pages in order, or jump to what you need:

1. Provision a GPU — fund your wallet, pick a GPU, and launch an instance.
2. SSH access — register an SSH key and connect from macOS, Linux, or Windows (with video walkthroughs).
3. Move data — upload datasets and pull results back with cp, send, and receive.
4. Fine-tuning — run managed podstack train LoRA/QLoRA jobs.
5. Pricing & billing — how per-hour pricing, spot vs on-demand, and the wallet work.
6. Troubleshooting — fixes for the common provisioning, SSH, and data-transfer issues.

Prerequisites

• A Podstack account with wallet balance — see Pricing & billing.
• The podstack CLI installed and authenticated — see CLI installation and authentication.
• ssh and scp available on your machine (standard on macOS/Linux; on Windows use PowerShell or Git Bash).

# Confirm the CLI is installed and you're signed in
podstack version
podstack auth whoami

FAQs

Which GPUs can I rent? Whatever is in the catalog — H100, A100, L40S, and more. List live options and rates with podstack gpu types list and podstack gpu pricing.

What’s the difference between spot and on_demand? spot is cheaper but can be reclaimed when capacity is needed; on_demand costs more but is more stable for long runs. Pick the mode when you create the instance.

How do I connect to an instance? podstack gpu instances ssh <id> opens a shell using your local SSH key. Register keys first with podstack gpu keys create. See SSH access.

Where does my SSH private key go? Nowhere — Podstack only registers your public key. The private key never leaves your machine.

How am I billed? Per hour, from your wallet, only while the instance is running. Delete the instance to stop billing. See Pricing & billing.

Can I provision from the dashboard instead of the CLI? Today, provisioning is through the podstack CLI. An in-portal GPU marketplace is built but not yet generally available.

When should I use managed fine-tuning instead of a raw box? Use podstack train when you want a LoRA/QLoRA model without managing a GPU. Rent a raw TrainPod when you need a custom framework, data pipeline, or full control of the machine. See Fine-tuning.