title: the three places where eval can run, and the limits of the fleet
source: manuals/miles-101/sections/6-2-run-a-separate-eval-fleet.md
lines: 13-19
lang: text
note: Read the cost column, then the two limits after the table.
---
| Where | How to select it | Cost |
|---|---|---|
| The training engines | the default | rollout production stops while the eval runs |
| A dedicated fleet | `--eval-num-gpus N` | N GPUs of the job go to eval, and training never stops for it |
| Your own service | `--eval-function-path` set to a checkpoint eval function | only the checkpoint export |

You cannot select the fleet together with a checkpoint eval function for your own service. Rollout fault tolerance from [section 6.3](#s-survive-failures) also does not work with a fleet.
