gateway: POST /captions/batch — N images per request, fanned out to vLLM

Retune for L40S (48GB): CUDA graphs back on, roomier memory split

No CPU fallback for the face model: refuse to serve without CUDA

Stability + persistence: OOM headroom, supervised vLLM, bucket cache/logs, open root

Install onnxruntime-gpu LAST (insightface drags in the CPU wheel)

Fix CUDA provider load for onnxruntime-gpu

vLLM JoyCaption (OpenAI-compatible) + GPU InsightFace faces endpoint behind one authed gateway

initial commit