Existing bare-metal GPUs
InferCrane does not install a node agent or take over a bare-metal scheduler. It can perform a
bounded, read-only inventory of the machine where the CLI runs:
Machine-readable output is available for setup scripts:
The v1 probe invokes nvidia-smi directly without a shell and records concrete GPU index, UUID,
model name, total memory, and driver version. It has a five-second deadline, bounded output, and a
256-GPU limit. It does not inspect prompt data, install software, open a port, change GPU state, or
send inventory to the control plane.
Hardware discovery is not compatibility or performance evidence. GPU presence does not prove
model fit, runtime/CUDA compatibility, network reachability, health, cost, or a production SLO.
AMD/ROCm inventory and remote host discovery are not implemented by this contract.
Connect the runtime already serving the host
Start or retain vLLM, SGLang, LiteLLM, or another OpenAI-compatible runtime under your existing
process manager. Then connect the service without transferring lifecycle ownership:
The connection begins observe-only. InferCrane may discover the served model and record health,
but it does not SSH to the host, restart the process, change GPU allocation, scale replicas, or
delete the external service. Explicitly move to traffic-managed only after reviewing the route and
credential boundary.
Use a lifecycle provider adapter only when you want InferCrane to create and delete capacity. A
future bare-metal lifecycle adapter would require a qualified external scheduler or agent and a
separate ownership-transfer design; read-only discovery does not imply that capability.