Thousands of files, one session.

Bring a script and its inputs, run it on a datacenter GPU, take the outputs back. The machine is deleted at the end.

For whom

• Embedding a corpus, transcribing audio, captioning images.
• Running evaluation suites and benchmarks.
• Any GPU job with a clear start and end.

Which GPU

Size by the largest model the job loads, plus batch memory:

Job	Typical VRAM
Embeddings, small encoders	16 GB
Speech-to-text (Whisper large)	16 to 24 GB
Vision-language captioning (7B)	24 to 48 GB
LLM evaluation (7B to 70B)	See LLM inference

Billing is per minute: a throughput-bound job on a faster GPU finishes sooner and stops billing sooner.

Steps

1. Book a machine. Choose a duration that covers the job plus a margin.
2. Upload code and inputs with rsync.
3. Start the job in a detached tmux session. It keeps running when you disconnect.
4. Follow the log over SSH.
5. Download the outputs, then end the session.

Tips

• Put the environment setup in the script (uv installs Python packages in seconds).
• Write outputs incrementally and make the job resumable. On interruptible offers, a reclaimed machine ends the session.
• Large inputs: download them on the machine directly from your own storage instead of uploading from your computer.
• Ending as soon as the job finishes returns the unused hold to your credit.