Qwen 2.5 VL OCR trained on ROCm

The AMD AI Academy Optical Character Recognition (OCR) challenge focuses on building an end-to-end vision pipeline capable of extracting text from complex road scene images under real-world conditions. Designed for autonomous driving and surveillance applications, the system recognizes text across diverse categories: US license plates (extracting registration numbers while filtering out state banners and slogans), Chinese license plates (retaining province prefixes), speed limit postings, advisory speed plaques, stop signs, and multi-line work zone signage. The pipeline is built to maintain high accuracy despite adverse environmental factors like glare, motion blur, sensor noise, low lighting, and severe perspective angles. Architecturally, the application leverages advanced open-source Vision-Language Models (such as Qwen2-VL) running on AMD ROCm GPU acceleration. To strictly meet evaluation constraints—including a 30-second per-image inference limit and a total 10-minute run budget—the application uses a persistent background daemon (FastAPI/vLLM). Model weights are loaded into VRAM once at container boot, allowing the CLI entrypoint script (app/app.py) to execute inference calls with sub-second latency via local HTTP requests. The entire application is packaged inside a Docker container built strictly on the mandated base image (rocm/pytorch:rocm10.0_ubuntu26.04_py3.14_pytorch_release_2.13.0), keeping the uncompressed image under 60 GiB and peak GPU VRAM usage between 1 GiB and 48 GiB. Custom post-processing logic enforces domain-specific rules before generating normalized output JSON files for automated scoring