ROCm Hyperloom
A multi-agent harness that autonomously tunes LLM and diffusion inference serving configurations, frameworks and GPU kernels for AMD Instinct GPUs, then stores proven results in a reusable recipe knowledge base.
ROCm Hyperloom is a multi-agent system that optimizes inference serving on AMD Instinct GPUs without per-model human tuning. It profiles a workload, searches serving framework and GPU kernel optimizations across vLLM, SGLang and xDiT, validates every candidate change end to end, and supports text generation, image generation and custom pipelines.
A session runs a closed loop of prelude, enablement, framework optimization, kernel optimization and sweep phases: it measures a stock baseline, optionally replays the closest known recipe, makes the model runnable if a baseline fails, searches serving flags, precision and batching choices, delegates hot kernels to one of two AMD kernel backends, and re-measures every accepted change. A hard cycle ceiling and a roofline-based headroom check decide whether another optimization cycle runs, keeping the agent grounded across long sessions instead of drifting or repeating dead ends.