[CrypticSwarm]

A tong could not ask for GPU access, so a model server like Ollama could not run as a tong with the GPU support make run_ollama gives it today (--gpus=all). This adds resources.gpus, passed to docker run --gpus, so that tong can be built outside this repo. The existing Ollama make targets are unchanged.

What changes

• resources.gpus: a new field, parsed by one helper (parse_gpus in swarmforge/tongs/model.py) that both validation and argv assembly call, so the two cannot disagree. Only these forms are accepted:

  • all
  • a positive count (an int or ASCII digits, at most 2³¹−1)
  • device=<id>, or the docker-quoted list "device=<id>,<id>", where each id starts with a letter or digit and continues with letters, digits, and ._:-

  Docker's other --gpus keys are refused. driver=, capabilities= and options= can grant more than GPU access (driver=cdi injects any device registered on the host). count= is redundant with a plain count. Leaving --gpus free-form would also have let control characters reach docker. An unquoted device=0,1, the likeliest mistake given docker's CSV parsing, is refused with the exact YAML quoting to use instead.

• Approval gate: GPU access is a host-device grant, so the privilege summary reports it and a workspace tong's approval prompt shows a gpus: <value> (host GPU access) line. The config hash already covers the whole definition, so adding or changing gpus asks for approval again and recreates a running shared tong. Tests cover both.

• Docs: a new Resources section in docs/tongs/definitions.md covers the accepted forms, the quoting, why the other keys are refused, and the host requirement. GPU containers need the NVIDIA Container Toolkit on Linux; Docker Desktop's WSL 2 backend provides GPU support itself. Without it docker run fails and the launch stops. docs/tongs/approval.md lists GPU access among what the prompt shows.

Testing

• make test: 948 tests pass (923 on master). Each commit also passes on its own.
• ruff check is clean.
• New tests cover every accepted form and every refused class (control characters, trailing newlines, bools, zero, over the cap, non-ASCII digits, the refused --gpus keys), plus the --gpus flag placement before the image in the tong's docker run argv and through the orchestrator, the approval prompt line, and a non-mapping resources value.
• Not exercised here: a real docker run --gpus against a GPU host. That needs a host with NVIDIA tooling. The quickest manual check is a user-layer shared tong with resources: {gpus: all} running nvidia-smi.