A machine requirement cannot say how much GPU memory it needs, so Stella starts on a card too small #23

Priority

extreme

Components

• app/nullius (backend)
• app/web (routes)
• templates and static files
• docs and the site
• catalogue (benchmarks, agents, capabilities)
• the standard (ports, kits, relay, runs)
• CI, workflows and skills

Today

A benchmark's machine requirement (host in its standard.json) is checked by programs on PATH only. LongMemEval's Stella retriever needs a CUDA GPU with more memory than an 11 GB card has, and says so only in prose: "Stella needs more than 11 GB of GPU memory." On a machine with an 11 GB card the builder shows the requirement as met (nvidia-smi is there), Run starts, and the unit fails later inside run_retrieval.py when its fixed batch of 64 runs out of memory.

Wanted

A requirement may say the least memory its GPU must have (gpu_memory_mb). The builder's check on this machine and the run itself read the largest GPU's memory from nvidia-smi and refuse a card that is too small, saying what is needed and what is there. Stella's two retriever values ask for it.

Acceptance

• host.<name>.gpu_memory_mb is validated as a positive whole number of MiB.
• A run whose options need such a requirement is refused when the largest GPU has less, naming the need and what the machine has.
• The builder's check for this machine shows the requirement as missing with the same words.
• A machine whose GPU cannot be asked is reported by its missing program, not as too small.
• LongMemEval's flat-stella-session and flat-stella-turn need a GPU of at least 12000 MiB; Contriever's values are unchanged.
• docs/standard.md says the field.

Dependencies and related

A host's GPU memory is not probed before Run; the run checks it as it starts on the host.