ROCm™ Hyperloom is an autonomous agentic system designed to optimize end-to-end inference workloads (targeting both host code and GPU kernels) on AMD GPUs. Using advanced AI agents and profiling tools, Hyperloom analyzes your workload, identifies performance bottlenecks, implements targeted optimizations, and validates the performance and correctness of the optimizations without requiring manual intervention.
The system operates through a sophisticated multi-stage pipeline. First TraceLens, the profiling brain of the workload understanding stage, consumes traces collected by Magpie (which in turn relies on IntelliKit for some low-level GPU profiling tools), captures bottlenecks, and derives the roofline targets that seed the optimization search tree.
Next, Hyperloom employs a self-evolving code optimization engine following an iterative agentic loop (Think → Decide → Implement → Benchmark). Arbor intelligently explores the optimization space using a Dynamic Specialist Agent and Knowledge Base. In parallel to Arbor, GEAK, a multi-agent GPU performance optimizer, optimizes hot kernels. Once optimizations are identified and validated, Hyperloom prepares the optimized code and generates a report with all proposed changes and expected performance improvements. This end-to-end automation enables developers to achieve significant performance improvements while maintaining code quality and reducing the manual effort traditionally required for GPU optimization.
Hyperloom combines:
- Trace analysis, identifying bottleneck kernels and bridge planning through TraceLens Agent (backend support from Magpie and Intellikit)
- Kernel optimization through the GEAK backend.
- Agentic search space exploration through Arbor, a tree-based cognition layer with dynamic agents, long-horizon campaigns, and self-evolving optimization guided by a curated knowledge base of hardware learnings, pitfalls, and prior campaign artifacts.
| Goal | Guide |
|---|---|
| Set up Hyperloom and run a demo | Quickstart |
| Launch and monitor an optimization | Run an optimization |
| Understand the algorithm | Optimization loop |
| Topic | Link |
|---|---|
| ROCm Docs | Hyperloom |
| Authentication and credentials | Authentication & credentials |
| Environment variables | Environment variables |
| Components | Components |
| Compatibility | Compatibility matrix |
| Troubleshooting | Troubleshooting |
| Operations | Operations & self-hosting |
| Session output schema | session_breakdown.json |
If you encounter any problem or bugs while running Hyperloom, feel free to open an issue, or provide us with feedback on how to improve Hyperloom by completing the beta survey.
- Runtime package:
src/hyperloom/ - Main agent instructions:
src/hyperloom/inference_optimizer/SKILL.md - CLI entry point:
python -m hyperloom.inference_optimizer.cli optimize - Operator tools:
python -m hyperloom.inference_optimizer.tools.* - Documentation source:
docs/
For contribution workflow, testing, and linting, see
CONTRIBUTING.md.
Hyperloom is released under the MIT License. The full license text
is in LICENSE.
You may use Hyperloom commercially, modify it, and distribute it under the terms of the MIT license, provided the copyright notice and the permission notice are retained in all copies or substantial portions of the software.
Third-party tools and agents (Cursor, Visual Studio, and Claude Code)
that Hyperloom invokes are governed by their own separate license terms
and are NOT covered by the MIT license above — see the "Third-Party
Tools and Agents" section in LICENSE. You are responsible
for reviewing and complying with each tool's individual license.
For security-relevant issues, see SECURITY.md. For
contribution conventions, see CONTRIBUTING.md.