# Live Inference Hardware Procedure

This procedure applies before any live local-model inference, benchmark,
extraction harness, model load, or token-generating request.

## Non-negotiable rule

Never assume that an endpoint name, model tag, service configuration, or an
inherited device variable selected the correct accelerator. Verify actual
runtime placement first.

## Required preflight

Before generating tokens:

1. Confirm the endpoint the user designated.
2. Confirm the exact model tag to run.
3. Inspect accelerator inventory and actual model-runner placement using a
   runtime hardware observation tool.
4. Confirm the runner is on the requested capable accelerator.
5. Confirm no runner for this test remains on an unintended or low-capability
   accelerator.
6. If an unintended runner is resident, unload it and verify that device is
   clear before proceeding.
7. State the preflight result in a working update.

If the requested accelerator cannot be identified or verified, stop and ask
for the endpoint/device mapping. Never silently fall back to a default endpoint,
CPU, consumer card, or lower-capability card.

## Zero-token exception

A zero-token model-load probe is allowed only after verifying that the server's
device visibility is restricted to the intended capable accelerator(s). Inspect
actual placement after the probe and before any request that generates tokens.

## Harness record

Persist or report:

```text
endpoint
model tag
requested accelerator
observed runner process/device
observed available memory/capability as relevant
time of preflight
```

Do not treat a successful HTTP response as hardware proof.

## Resource behavior

- Do not start broad parallel model work without checking device capacity.
- Pin managed launches to an approved device identity rather than relying on
  device index order.
- Do not stop or unload an unrelated model/process.
- When a test ends, preserve user work and leave runtime state as requested.