🤓 Latest Submissions
ROCmigrator / Autonomous CUDA-to-ROCm Agent
ROCmigrator is an autonomous engineering agent that transforms PyTorch CUDA C++ extensions into verified AMD ROCm/HIP implementations. It addresses a critical challenge in GPU software portability: translating CUDA code is only the beginning. A reliable migration must also compile, execute correctly, preserve numerical accuracy, and produce reviewable changes. ROCmigrator implements an explicit 12-stage state machine covering repository acquisition, security scanning, CUDA analysis, CPU reference generation, deterministic HIPIFY conversion, real HIP compilation, GPU verification, benchmarking, reporting, human approval, and GitHub pull request creation. The system combines a Next.js frontend, FastAPI backend, Pydantic-validated contracts, isolated Docker execution, and a bounded AI-assisted repair mechanism. Deterministic tools are prioritized; AI repair is invoked only when necessary and remains subject to strict limits. We validated ROCmigrator on a real AMD Instinct MI300X VF (gfx942) using AMD Developer Cloud. The Golden 01 CUDA extension was successfully transformed, compiled with the ROCm HIP toolchain, and executed on AMD hardware. All 18 numerical comparisons passed across three random seeds, three tensor shapes, and float32/float64 data types. CPU reference verification and determinism checks also passed. The successful run required zero AI repair rounds. After verification, ROCmigrator paused for explicit human approval. Only after approval did the agent create GitHub Pull Request #1 containing the migrated changes and verification evidence. ROCmigrator demonstrates a complete CUDA-to-ROCm engineering workflow: from source transformation to real GPU execution, numerical validation, and human-controlled delivery.
18 Oct 2026
.png&w=828&q=75)
MC3: Hardware-Optimized RAG Pipeline for AMD GPU
This project presents a highly optimized, fully containerized Retrieval-Augmented Generation (RAG) pipeline developed for Mini Challenge 3 of the AMD AI Academy Challenge. Built strictly upon the mandated rocm/pytorch base image, the solution processes proprietary domain knowledge—including multi-modal documents, spreadsheets, and source code—to deliver accurate answers paired with exact file citations. Designed to strictly adhere to the zero-hallucination policy, the architecture ensures that responses are grounded entirely in the provided corpus. The system is engineered for seamless deployment on AMD hardware, utilizing /dev/kfd and /dev/dri device mappings for native GPU acceleration. By restructuring the pipeline into a persistent long-running process, it completely eliminates the startup overhead per question, ensuring the system easily fits within the stringent 30-second query budget. Despite massive physical constraints, the final environment was successfully packaged into a unified image and pushed to Docker Hub. This solution demonstrates robust handling of complex dependencies, encrypted files, and unknown data types, ensuring high-throughput hardware-accelerated inference in AMD Instinct and Radeon environments.
1 Dec 2026
