Hyperloom

Hyperloom is an autonomous agentic system from AMD, part of the ROCm platform, built to optimize end to end inference workloads on AMD GPUs. It uses AI agents together with profiling tools to analyze a workload, find performance bottlenecks, implement targeted optimizations, and then validate both the speed and the correctness of the result, without requiring manual intervention at each step. It targets engineers and teams who serve large language models on AMD hardware and want to shorten the tuning cycle. The documentation covers installation on Docker, bare metal or a Slurm cluster, the optimization loop and its components, running an optimization on a supplied workload or on your own, quantization with AMD Quark, kernel optimization, multi node inference optimization, a knowledge base integration, operator scripts, an API reference and a session output schema. Hyperloom runs on Linux and obtains model access through an API key from the user's organization LLM gateway. Hyperloom is released under the MIT license and its source code is hosted on GitHub under AMD AGI. No paid plans are published, since it is distributed as open source software, although running it requires access to an LLM gateway and GPU hardware that carry their own costs. The documentation describes version 1.0.0, dated August 2026.