OCR engine for RLX — native compiled detection + recognition graphs.
Load HuggingFace robertknight/ocrs weights as
.safetensors (use rlx-ocr-convert with feature convert-rten to export from legacy .rten).
Detection + recognition run on every standard RLX backend when the matching rlx-runtime
feature is enabled (cpu, metal, mlx, cuda, rocm, gpu, vulkan). Build with
all-backends for a single binary that accepts all --device values.
Parity / baseline: parity-ocrs (upstream ocrs + RTen) or rten-inference (RTen graphs only).