"Cubiczan team on AMD AI League Hackathon"

Team Idea

We're building ROCm Migration Copilot: an open-source AI agent that ports CUDA code to HIP/ROCm. Point it at a repo: it translates kernels, compiles in a ROCm sandbox, verifies results vs the CUDA baseline, benchmarks on AMD Instinct MI300X. No vendor lock-in, no weekend-long porting pain. Stack: Python, LLM function-calling loops, ROCm 6.x, PyTorch, vLLM, Docker, React, on AMD Developer Cloud credits. I'm the lead AI/ML engineer driving architecture and roadmap, seeking LLM/ML engineers to own modules: agent loop, CUDA-to-HIP knowledge base, parity harness, benchmark tooling. Motivation beats experience; if GPUs, agents, or AMD's ecosystem excite you, you'll learn fast.

Team Members

Shyam Desigan

impactquadrant655

Submission

ROCm Migration Copilot: AI CUDA-to-HIP Agent

An AI agent that ports CUDA codebases to AMD HIP/ROCm automatically — analyzing, translating, building, and tuning GPU code so any project can escape CUDA lock-in.