Keep the context of a project together and reuse it when requesting compute.
Create a workspace
Open Workspaces, choose a project name, describe the workload and select its type. Add an existing S3 location if useful, then select Create workspace.
A workspace saves configuration and data references. It does not create a bucket, upload files or start a GPU session. Workload types describe your requirements; they do not install a framework or container image.
Request compute
Open your workspace and select Get compute for this project. The request includes your saved objective, workload type and data reference. Review those details, then choose the GPU count, location, UTC window and budget.
Submitting a request does not charge your account or reserve GPUs. You can find submitted requests again from the workspace.
Update the project
Edit the configuration and select Save changes. Existing requests keep the workload details you originally submitted. If another edit happened first, reload the current configuration before saving again.
Data and access
Workspaces are private to your signed-in account. Enter S3 references such as s3://your-bucket/project/, without credentials or signed URLs. Saving a reference grants no access to the bucket. Keep API keys, passwords and sensitive data out of free-text descriptions.
Run code on your reserved GPUs
When a matching offer supports Workspace · Python and shell, choose that access mode before accepting the quote. Once the reservation is ready, open My workspace from its reservation page.
- Select Check GPU access. The check reports the GPUs visible inside your environment.
- During your reserved window, select Confirm access & start session. This starts the usage counter; preparing the environment does not.
- Choose Python or shell, enter your code and select Run on my GPUs. Read its exit status and output under Command results.
- Select Stop session & release GPUs, or tick End the reservation when this command finishes before running code. The reservation shows the final charge and returned credits after it ends.
One command runs at a time, for up to five minutes and within the reservation window. Code is limited to 4 KiB and displayed output is bounded. The supplied Python environment includes PyTorch; network access follows the approved supplier policy. This interface executes commands in your GPU container; it does not grant administration of the supplier's host.
Files in /workspace last only for that GPU reservation. Save results before stopping. Your project description and data references remain in your account, but this does not persist the files from the GPU environment or mount an S3 bucket automatically.
Connect with SSH or VS Code
For a running session with an SSH-capable supplier, open SSH & VS Code on the session page. Paste your Ed25519 public key from ~/.ssh/id_ed25519.pub, then download the connection file. Never upload your private key.
Download the SSH helper from that page and run the displayed setup command. It requires Python 3.10+ and OpenSSH on macOS or Linux. The helper creates a dedicated Python environment, stores the temporary credential with restricted permissions and adds a Requenta host to your SSH configuration. Existing SSH entries are preserved. Use the printed ssh command or choose that host in the VS Code Remote-SSH extension. Open /workspace to work with the same files as your browser commands. SFTP is available to copy results to your computer.
Access lasts up to one hour, within your reservation. Creating a new connection file replaces the previous access. Revoke access closes tunnels within 15 seconds; stopping the reservation also closes access. Reconnect using a fresh file when access expires. Delete the downloaded JSON after setup; it contains a temporary connection credential.
The helper checks the SSH host key supplied by your session. Your computer and the supplier agent connect outward to Requenta over HTTPS; the supplier does not expose an incoming SSH port. SSH gives access to the allocated container, without host administration. Only loopback TCP forwarding is supported. VS Code can download its server locally and transfer it over SSH when the supplier environment cannot reach the Internet; its extensions remain subject to the supplier network policy.
SSH availability is confirmed by the supplier agent. If setup is unavailable, the existing browser command interface remains accessible for a running workspace. Storage is still temporary and is not shared automatically between reservations.
Disk and transfer budget
Choose working-disk capacity from the supplier’s available sizes before accepting the quote. Additional disk appears as a separate price line and is billed for delivered minutes. /workspace remains temporary: its files are deleted during cleanup. Save important results before stopping. Exceeding the Kubernetes disk limit can evict the workload.
New resource bundles have separate upload and download allowances, shared across SSH, SFTP, VS Code and loopback forwarding. Check both counters before transferring large models. They reserve blocks up to 1 MiB and include SSH overhead and unused tails. When either budget runs out, SSH cannot operate normally; you can purchase an explicit extension, or stop compute from the console. Compute keeps running until stopped or expired.
The reservation shows optional upload/download prices before acceptance. Extensions require a separate confirmation and sufficient credits. Unused whole GiB return to credits at session end. Earlier accepted profiles retain their original 10 GiB disk and combined transfer budget. See Pricing.
Download from approved destinations
Some suppliers offer controlled downloads. On those offers, choose what your session may download from before requesting the quote: Hugging Face, PyPI, GitHub, a regional S3 endpoint and up to five exact hostnames. The choice is part of the quote and cannot change during the reservation. Everything else stays blocked.
Your session receives HTTPS_PROXY automatically. pip, huggingface_hub, boto3, git and most HTTPS clients use it without configuration. Only HTTPS on port 443 works; plain HTTP and other ports are refused. Downloads count in your transfer-to-GPU allowance and uploads in your transfer-from-GPU allowance, at the rates shown on the quote. When an allowance runs out, new connections are refused while compute continues; you can buy an extension or stop the session.
Byte counts for these downloads are reported by the supplier's execution agent, as compute time is. SSH and SFTP transfers are counted by Requenta's relay. To read a private bucket, bring its credentials into the session yourself, for example as environment variables in a command. Requenta does not store them. Bucket charges remain with your bucket provider.
Use data from a personal bucket
A saved S3 location is a project reference, not a bucket connection. Your bucket remains on your own cloud account. Its storage, API, retrieval and outgoing-data fees are billed by that provider and are outside Requenta’s quote.
Without controlled downloads, fetch the required files to your computer using your existing S3 client, then copy them into /workspace with SFTP using the SSH host created by the Requenta helper. Copy checkpoints back before stopping, then upload them to your bucket. Bucket credentials stay on your computer. The Requenta upload/download budgets cover the SSH part of this path.
For large datasets, rclone can stream between an existing S3 remote and the Requenta SFTP connection through your computer, without staging the whole dataset on its disk. Configure your own S3 remote first, then use the SSH host printed by the helper:
rclone copy my-s3:my-bucket/models :sftp:/workspace/models --sftp-ssh "ssh requenta-BOOKING-ID" --transfers 1 --checkers 1 --sftp-disable-hashcheck
Replace my-s3, the bucket path and requenta-BOOKING-ID with your configured remote and actual SSH host. For results, reverse source and destination. Keep the computer connected, ensure the session’s disk and both transfer allowances are sufficient, and renew SSH access if it expires. This is a client-mediated copy, not a persistent mount or a direct GPU-to-S3 network connection. See the rclone SFTP documentation for external SSH configuration.