PlateParse: ROCm Vision OCR for Plates and Signage

PlateParse is a vision-language OCR system for vehicle registration plates and road signage, built to run entirely on AMD Instinct GPUs via ROCm. It uses Qwen2.5-VL-7B-Instruct as the recognition backbone, served from a container whose final stage is the mandated rocm/pytorch ROCm 10.0 base, with model weights staged in a separate build stage so they never bloat the base image. The scoring pipeline separates recognition from transcription rules. The vision model does the reading; normalize.py then applies category-aware cleanup: Chinese license plates keep their province prefix, single-token alphanumeric strings are treated as plate numbers so state banners and mottos are stripped, and multi-token output is treated as road signage where all text is preserved. Units appended to advisory speed plaques are removed without damaging genuine sign text. Both sides of grading are normalized identically, so emitting the correct character set is what matters, not formatting. The system was debugged directly on an AMD Instinct MI300X under ROCm, which surfaced a hardware-specific failure worth noting: the Qwen2.5-VL vision tower segfaults under bfloat16 on this ROCm generation. Pinning float32 weights with eager attention resolved it. Verified end to end on the MI300X, reading a synthetic California plate as 7ABC123 at 0.9988 confidence, with the banner correctly stripped. The container is defensive by design: it always writes schema-valid JSON and exits zero, so a single malformed image costs its own points rather than ending the run and zeroing every remaining image.