Samiullah Hafeez@samiUllah

This project implements an image-to-text OCR pipeline using the Qwen3.5-4B vision-language model on AMD GPU hardware. It processes an input image and extracts visible text, returning the result as structured JSON. The application uses PyTorch and Hugging Face Transformers to load the model and processor. The model is loaded once and reused for subsequent image requests to reduce repeated initialization overhead. The application accepts an image path through the command-line interface and writes the extracted text to an output JSON file. The solution is packaged as a Docker container based on the required ROCm PyTorch image, supporting execution in the challenge environment. The container includes the application code and its dependencies, and the model is available locally in the tested image. This approach provides a reproducible deployment workflow for GPU-based OCR. The project focuses on practical deployment, structured output, model reuse, and compatibility with the challenge's containerized evaluation process.

Six Matches. One leaderboard. Build on AMD, earn points and compete for prizes.