ProfiloAI - AMD GPU Doctor

ProfiloAI is an AMD GPU performance assistant for ROCm developers. It helps users turn profiler output from rocprof, omniperf, or training-loop metrics into practical engineering advice. Instead of only showing raw numbers, ProfiloAI explains the likely bottleneck, the root cause, a concrete code-level fix, and an expected speedup range. The project was built for the Fine-Tuning on AMD GPUs track. The training and benchmark run were completed on AMD Developer Cloud using an AMD Instinct MI300X GPU, ROCm, PyTorch, Hugging Face Transformers, PEFT LoRA SFT, and TRL DPO alignment. The benchmark improved from a 40.5% base model score to a 46.5% ProfiloAI score, a +14.8% relative improvement. The public Hugging Face Space runs in lightweight demo mode so judges can test the workflow without needing a live GPU server. The GitHub repository includes the full app, dataset generation scripts, AMD Cloud runbook, training scripts, evaluation scripts, benchmark comparison, and optional vLLM-compatible serving path.