BOSU MESIPAMU@Bosumesipamu

Here's a description in the 1700-1800 character range — comfortably inside 600-2000: --- **GPU Goblin Auto-Tune** is an AI-driven optimizer that finds the fastest configuration for any HuggingFace fine-tuning workload on AMD MI300X (ROCm 7.0, CDNA3, 192 GB HBM3). Point it at a model id and walk away — it returns a measured speedup and a ready-to-ship modified training script in roughly ten minutes. **The problem.** Manually tuning LLM fine-tunes on MI300X is brutal: dozens of interacting knobs (precision, batch size, attention implementation, dataloader, optimizer, env vars), ROCm-specific gotchas distinct from CUDA, and "best practices" that sometimes hurt on your specific workload. Generic hyperparameter tuners don't know MI300X. **The solution.** A single CLI that profiles a baseline, iteratively tries curated MI300X-specific changes (bf16, larger batches, SDPA attention, hipBLASLt, TunableOp, MIOpen FAST), benchmarks each on real hardware, and keeps what wins. Three modes: hardcoded playbook (deterministic, no API key), LLM-driven greedy (Qwen picks one change per iteration based on live waste-budget signals), and LLM-explore (Qwen proposes K candidates; the script benchmarks all and tries a merged version of the positives to capture compound gains). **Output.** A live per-iteration log plus a final report showing tokens/sec delta, MFU jump, waste-budget reduction chart, accepted vs rejected experiments, and a downloadable best.py. **Real result on Qwen2.5-7B LoRA fine-tune:** baseline 5,734 tok/sec → tuned 11,708 tok/sec (2.05×, +105%; MFU 20% → 41%) in ~10 minutes, fully automated. **Architecture.** Streamlit UI deployed to a Hugging Face Space; FastAPI server on the MI300X droplet streams NDJSON events back to the browser as SSE. Judges click the Space, enter a model id, and the GPU work happens on AMD hardware while progress streams live. Powered by Qwen as both the agent brain and the demo audit target.

Grab your ideas, spin up compute, and start building what’s next! Start with $100 in AMD Developer Cloud credits and begin building right away. Get access to hands-on tutorials, workshops, and curated learning resources to help you build faster and go further. To unlock credits and join the hackathon, sign up for the AMD AI Developer Program. 🤝 Go solo or team up — ship something real. 💰 $21,500+ total prize pool & 1 × AMD Radeon™ AI PRO R9700 GPU up for grabs. 📍 On-site Venue (May 9-10): MindsDB SF AI Collective 3154 17th St, San Francisco, California, USA 📍On-site participation is by invitation only. Travel and accommodation expenses will not be covered. 🧑‍💻 Developers, builders, and innovators - bring your boldest ideas and let’s turn them into real AI systems on AMD.