Capacity Planner
Enterprise module
The Planner is part of the GPU Fleet module (gpu-fleet) of your Cognipeer package.
The Planner tab of Operate → GPU Fleet answers one question: which models can run together on this machine, and how much memory is left? You can plan on a machine that is already in your fleet or on one you are thinking of buying.
Choose a machine
Under Machine, pick a source:
- In the fleet: one of your online machines. Its GPUs, MIG slices and live memory use are taken from what the agent reports. If no machine is online yet, plan on one to buy.
- To buy: pick a GPU or chip and how many GPUs. The list is grouped as NVIDIA · Data center, NVIDIA · Workstation, NVIDIA · GeForce and Apple Silicon.
- For NVIDIA GPUs that support MIG, Partitioning lets you plan on Whole GPUs or on MIG slices (for example
3× 2g.20gb). - For Apple Silicon, pick the Unified memory size. macOS lets the GPU use only part of it by default, and MLX models run inside that share.
- For NVIDIA GPUs that support MIG, Partitioning lets you plan on Whole GPUs or on MIG slices (for example
From a machine's page, Plan capacity in its menu opens the Planner with that machine selected.
Add models
Under Models to run together, add the models you want on the machine. Search Hugging Face, or pick one of the models Suggested for this machine. For each model you see the memory it needs (minimum to recommended), and you can choose its Runtime and the number of Copies. A model that no runtime can start on the machine says No runtime starts here.
The header shows how many of the models fit (for example "2 of 3 fit").
Read the result
For every model the Planner shows where it is placed: the GPU or MIG slice and the memory it holds, or Fits nowhere here. The Memory summary shows four numbers:
| Number | Meaning |
|---|---|
| Capacity | Total memory of the machine's GPUs (or the GPU share on a Mac). |
| In use | The live reading where the agent sends one, otherwise the estimates of the models already there. |
| Planned | What the models you added would take. |
| Free | What stays free, per GPU or slice (or how much a slice is over). |
Models of unknown size that already run on the machine are counted as "+n of unknown size".
Things to know
- Fit and speed are the same numbers that Run a model shows. They are estimates, not benchmarks; a measurement on identical hardware wins.
- vLLM reserves a share of its GPU and does not start without it, so every vLLM model, and every copy of one, needs a GPU or MIG slice of its own.
- On a Mac, every MLX model is a separate process with its own weights, all inside the GPU share. See Mac (Apple Silicon) and MLX.
- For a machine in the fleet, Run next to a placed model opens the Run a model flow.

