推論系統
AMD Open-Sources Hyperloom to Automatically Rewrite and Validate GPU Inference Optimizations Through a Multi-Agent Closed Loop
AMD has released ROCm Hyperloom 1.0.0a1, enabling agents to handle everything from performance profiling to kernel and host-code modifications, then decide whether to retain changes based on correctness and speed tests. The initial release supports only three Instinct GPUs, and the claim of reducing optimization time “from weeks to hours” remains AMD’s own; adopters will need to reproduce the results with their own models.
AMD has packaged an inference-tuning workflow that previously depended heavily on GPU experts into Hyperloom, an open-source multi-agent system. It uses a `Profile → Analyze → Plan → Optimize → Validate` closed loop: TraceLens-Agent first collects execution traces and identifies bottlenecks, then components such as Magpie and IntelliKit evaluate kernels and generate candidate modifications in HIP, Triton, or FlyDSL. Every change must pass correctness and performance checks before it is included in the final report. This goes beyond automatically searching launch parameters; it may modify GPU kernels, host code, and parallelization settings across repositories.
The first public release can be installed as a Python wheel, with a setup skill for Cursor, Claude Code, or Codex used to configure the working environment. Support is currently quite limited: Ubuntu 22.04/24.04, ROCm 7.2, and the MI300X, MI325X, and MI355X. Supported inference frameworks are limited to SGLang 0.5.12 or later and vLLM 0.21.0 or later. Engineering teams should also note that allowing agents to directly rewrite and execute low-level code means test nodes, model files, credentials, and source code should be placed in a disposable, restricted environment. The fact that the system validates results should not be treated as equivalent to comprehensive security isolation.
AMD claims Hyperloom can reduce end-to-end optimization time from weeks to hours, but there is not yet a comprehensive, independently comparable benchmark spanning different hardware and models. Public materials also do not break down agent invocation costs, search budgets, or failure rates. The next areas to watch are whether candidate patches can be reproduced easily, whether their gains transfer reliably across different batch sizes and context lengths, and whether support expands to Radeon GPUs, additional ROCm versions, and other inference engines.