Docs

Two docs. Same labels.
Different jobs.

The MIT repo is the plug: Docker starts, labels go on the container. Enterprise is the grant: A OOMs locally, B keeps serving.

Compose labels go to the MIT runtime. The enterprise daemon turns them into a grant.
Same compose file. The daemon is what makes vram.limit real.
MIT

Open source

Install vram and --runtime=vram. GPU nodes in the guest. Labels recorded. No daemon in that repo.
Commercial

Enterprise

Daemon, libvram, imported CUDA heap, dashboard. The two-Ollama shot: A at 512m dies, B at 4g still answers.