ROCm Migration Copilot

Most AI teams write code assuming NVIDIA CUDA — from Docker base images to torch.cuda device checks to NCCL for distributed training. When they want to evaluate AMD GPUs, they face a painful question: what exactly needs to change? ROCm Migration Copilot solves this problem. Users paste their CUDA code, Dockerfile, PyTorch training scripts, or logs into the app. The tool instantly detects NVIDIA-specific patterns with severity levels, generates a step-by-step ROCm migration plan, creates an AMD Developer Cloud benchmark checklist, and exports a complete Markdown report. The app is powered by Fireworks AI for intelligent analysis with a deterministic local fallback — ensuring the demo always works even without an API key. It is fully containerized with Docker and deployable on AMD Developer Cloud. Key features: - 7 pattern detectors covering CUDA runtime calls, custom kernels, NVIDIA container images, PyTorch CUDA assumptions, TensorFlow GPU checks, NCCL dependencies, and Triton kernels - Risk score from 0 to 100 with effort estimation - Step by step ROCm migration plan - AMD Developer Cloud benchmark checklist - One click Markdown report export - Fireworks AI powered analysis with local fallback Built with FastAPI, Python, Fireworks AI, Docker, and vanilla HTML CSS JavaScript frontend.