gpu-allocation-fallback

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

# GPU Allocation Fallback

Send typed results to `POST /api/agent/gpu-allocation/attempt` with a stable
`logical_allocation`, the requested compatibility/execution/disk invariants,
typed failure evidence, and the identical compatible preemptible candidate.
The grounded route uses zero model tokens.

The optional `success` field must be a literal JSON boolean. Omitting it records
the attempt as unsuccessful, the same as `"success": false`; only
`"success": true` resets the tracked allocation as succeeded. Strings, numbers,
`null`, arrays, and objects are rejected with HTTP 400 without changing fallback
or confirmation-token state.

Only quota/capacity exhaustion, insufficient GPU or Unschedulable, and no
compatible product/affinity count. For a Nebius Compute failure, send the
operation's `status.details[].code` as the failure code and the violation text
as the message: `NotEnoughResources` counts as capacity exhaustion, and
`QuotaFailure` counts only when the message names a `compute.instance.gpu.*`
quota. Auth, RBAC, network, image-pull, checkpoint, application, runtime,
cancellation, and timeout failures never count, even when a placement code is
also present. A timeout means the `timeout` code or a message that says the
operation timed out; the `NotEnoughResources` message's "VM schedule timeout"
wording is a capacity failure. The
default prompt occurs on the third qualifying failure, or immediately when
deterministic preflight proves on-demand cannot succeed and compatible
preemptible capacity is available.

Never switch automatically. Present the returned question and proposed action.
Accept through `POST /api/agent/gpu-allocation/consent` with its single-use,
action-digest-bound `confirm_token`; decline with `accept: false`. A decline
keeps on-demand and suppresses the same evidence until materially new evidence
arrives.

Preserve GPU family/product/count, image/digest, SM and RT-core requirements,
backend, model, workload tier, execution mode, and boot-disk count/bytes. The
state record contains only redacted digests, classification, selected pool, and
consent outcome. Success or a changed logical allocation resets attempt state.

Verify changes with:

```bash
npa/.venv/bin/python -m pytest \
  npa/tests/cli/test_agent_gpu_allocation_fallback.py \
  npa/tests/cli/test_agent_backend_render.py -q
```

From nebius/nebius-physical-ai, Apache-2.0.

More from nebius/nebius-physical-ai