- vllm/vllm-openai base: aligned torch/vllm/CUDA; onnxruntime-gpu rides the same CUDA 12 family
- buffalo_l baked into the image (build-time asserted), loaded on CUDA with provider reporting
- gateway: bearer auth (VALID_API_KEYS, fails closed, /health open), streaming /v1 proxy,
batched /faces/detect with queue_wait/gpu_ms/in-flight instrumentation, /stats percentiles
- uid-1000 user, local-disk HF_HOME with optional /data cache restore (patterns from
huggingface-engines/tts-dialogue-engine-moss)
