ROCm Hyperloom

Hyperloom is an autonomous agentic system that optimizes end-to-end inference workloads on AMD GPUs. It profiles your workload, pinpoints bottlenecks in both host code and GPU kernels, implements targeted fixes, and validates the results without manual intervention.

Traced through TraceLens, accelerated by GEAK kernel agents, and powered by Arbor tree-search, Hyperloom combines:

• Trace analysis — identifies bottleneck kernels and bridge planning via the TraceLens Agent (Magpie + IntelliKit backends).
• Kernel optimization — GPU compute hotspots auto-tuned via the GEAK backend.
• Agentic search — Arbor, a tree-based cognition layer with dynamic specialist agents, long-horizon campaigns, and a curated hardware knowledge base of pitfalls and prior campaign artifacts.

Supported GPUs

GPU Family	Models	Architecture	GFX Target	ROCm Support
Instinct MI300	MI300X, MI308X, MI325X	CDNA3	gfx940, gfx942	ROCm native
Instinct MI350 / MI200	MI355X, MI250X, MI200	CDNA4 / CDNA2	gfx950, gfx90a	ROCm native
Radeon RX 6000 series	RX 6900 XT, RX 6800 XT, RX 6800, RX 6700 XT, RX 6750 XT, RX 6650 XT, RX 6600 XT	RDNA2	gfx1030, gfx1031, gfx1032	Linux ROCm
Radeon RX 7000 series	RX 7900 XT, RX 7900 XTX, RX 7800 XT, RX 7700 XT	RDNA3	gfx1100, gfx1151	Linux ROCm
Radeon RX 9000 series	RX 9070 XT, RX 9070, RX 9060 XT, RX 9000 XT	RDNA4	gfx1201, gfx1206, gfx1207	Windows ROCm 7.3+ (WMMA FP8)

Representative peak FP16 (dense) TFLOPS for entry-level consumer cards:

GPU	CUs	Boost Clock	FP16 Dense	FP16 Achievable
RX 6600 XT	32	2.586 GHz	8.3	6.5
RX 6700 XT	40	2.42 GHz	24.8	18.0
RX 6800 XT	72	2.25 GHz	41.5	28.0
RX 9070 XT	64	2.66 GHz	194.8 (matrix)	170.0
RX 9060 XT	32	2.53 GHz	102.7	71.2

Get Started

Goal	Guide
Set up Hyperloom and run a demo	Quickstart
Launch and monitor an optimization	Run an optimization
Understand the algorithm	Optimization loop
Vulkan LLM engine (rdna4-llm)	Architecture

Toolkits

Hyperloom vendors or wraps several toolkits — run python aicompass.py status for a live summary of what is available in your environment.

Toolkit	Location	Role
Hyperloom	src/hyperloom/	Core orchestrator, CLI, agent loop
GEAK	vendor/GEAK/	Multi-agent GPU kernel optimizer
Magpie	vendor/Magpie/	Trace collection & benchmark runners
TraceLens	src/TraceLens-RDNA4/	Profiling analysis & roofline targets
IntelliKit	src/intellikit-RDNA4/	LLM engine, metrix profiling, agent-reach
Arbor	src/arboreal/	Tree-search cognitive layer
rdna4-llm	src/rdna4-llm/	Vulkan compute-shader LLM inference engine

Developer Entry Points

Purpose	Path
CLI entry point	python -m hyperloom.inference_optimizer.cli optimize
Main agent instructions	src/hyperloom/inference_optimizer/SKILL.md
Operator tools	python -m hyperloom.inference_optimizer.tools.*
Unified tool CLI	python aicompass.py status
GPU arch resolution	src/hyperloom/inference_optimizer/gpu_types.py
Roofline specs	src/hyperloom/orchestrator/kernel/roofline_ceiling.py
Per-platform arch JSONs	src/TraceLens-RDNA4/TraceLens/Agent/Analysis/utils/arch/

Documentation

Topic	Link
ROCm Docs	Hyperloom
Authentication and credentials	Authentication & credentials
Environment variables	Environment variables
Compatibility matrix	Compatibility matrix
Troubleshooting	Troubleshooting

Third-party / Vendored components

Path	What	Usage
third_party/vma/vk_mem_alloc.h	Vulkan Memory Allocator (header-only)	used by rdna4-llm buffer allocation
src/rdna4-llm/third_party/vma/	VMA copy for current engine	same
vendor/agentreach/	Agent Reach v1.5.0 (internet capability layer)	python aicompass.py agentreach or agent-reach

All vendored code is unmodified upstream, MIT-licensed. The AMD Radeon Developer Tool Suite is a manual download — gitignored, never synced to git.

---

Acknowledgments

This repo vendors Agent Reach (v1.5.0, MIT) to give AI agents internet access. Agent Reach is built and maintained by Panniantong (@Panniantong).

---

Licensing

Hyperloom is released under the MIT License. See LICENSE. Third-party tools and agents are governed by their own separate license terms.