Apppublic

arnoldkigozi0/Inquiline

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes

App README

Inquiline — GPU worker

Stateless GPU hands for an Inquiline instance running somewhere without a GPU. It holds no state: the orchestrator, ledger, tracer and skills all live on the always-on box, because a Space's disk is non-persistent and it sleeps when idle.

The hardware must be set to ZeroGPU in Space settings — this file cannot set it, and on any other hardware the @spaces.GPU decorators are inert.

Endpoints

Every input is a plain string so a stdlib HTTP client can drive it — no multipart, no FileData.

`api_name`	in	out
`imagine`	`prompt`, `steps` (1-4)	image
`see`	`image_url`, `question`	text
`transcribe`	`audio_url`	text
`health`	—	JSON: what loaded, what failed, is CUDA real

health carries no @spaces.GPU decorator, so probing it costs zero quota. It is the call to make first during bring-up, and the one that still answers when every model failed to load.

Driven from Inquiline by orchestrator/zerogpu.py; the names above are that contract.

Quota

Free HF accounts get 5 GPU-minutes per day and may host 2 ZeroGPU Spaces (verified email, account older than 30 days). The GPU is only consumed inside a decorated function, so idle time and the CPU-side URL fetches are free. An SDXL-Turbo image is 1-2 seconds.