Guides¶

Step-by-step walkthroughs for common FLAME workflows.

Training a simple Vision Transformer: a single-GPU benchmark. Train a ViT on mini-ImageNet (a public 100-class ImageNet subset) end to end, from "I have access" to a model training on a GPU. The same task is shown two ways — read them in order:

• Interactive (JupyterHub) — build it up step by step in a notebook and watch it train live. Start here.
• Batch (TrainJob) — save the training as a script and submit it to run unattended.

Other guides:

• Designing GPU workloads for preemption — checkpoint and resume so borrowed-GPU jobs survive eviction.