.png&w=1920&q=75)
PlateSign OCR on AMD ROCm
PlateSign OCR reads licence plates and road signs from uncooperative images on AMD GPUs. It runs Qwen3-VL-8B-Instruct in BF16 with Hugging Face Transformers on ROCm 10.0, packaged on the mandated rocm/pytorch base image. A resident server loads the model once at container start and warms up the GPU before it reports ready, so each grader call to app.py is a thin standard-library client that sends the image path over a Unix socket and writes the JSON answer. The prompt encodes the transcription rules: keep only the registration number on US plates (drop state names and slogans), keep the province character and letter on Chinese plates, keep every printed word on signs, and join multi-line text top to bottom. Tested in a grader-style run on an AMD Instinct MI300X: 10 of 10 sample images correct across PNG, JPEG and TIFF, 20 s startup, 0.16 to 0.35 s per image, 18.4 GB peak VRAM, 45.7 GiB uncompressed image.
Category tags:
For Innovators & Creators