GEAK

GEAK#

GEAK (Generating Efficient AI-Centric Kernels) is a multi-agent framework for end-to-end GPU kernel optimization in real codebases. It runs a closed loop of profiling, optimization, and validation, and produces reviewable patches backed by reproducible benchmarks. GEAK supports Triton, HIP (and CUDA / Composable Kernel (CK) / HSA Code Object (HSACO)), and FlyDSL kernels. It is driven by Claude Code and ships two deterministic JS Workflows — e2e_workflow for whole-model serving throughput and kernel_workflow for single kernels — with a deterministic control plane (budget loop, parallel fan-out, verification, stop conditions) that invokes LLM agents only for judgment.

Within Hyperloom, GEAK is the whole-pipeline end-to-end optimization delegate: when a workload is handed off, the orchestrator invokes GEAK once at the kernel-agent phase through the stable interface/run_e2e.py contract (a handoff.json in, a result.json back). Parallel exploration of candidate kernels then happens inside GEAK’s Workflows on the on-box GPUs.

Role in Hyperloom#

Hyperloom uses GEAK as the whole-pipeline e2e delegate when KERNEL_OPT_BACKEND_ORDER=geak (the bare-metal default). In this mode the orchestrator hands the optimization workload to src/hyperloom/agents/kernel/tools/backends/geak_runner.py, which resolves the GEAK checkout and launches GEAK’s e2e runner (interface/run_e2e.py) with the generated session context.

When GEAK owns the phase it runs the whole optimization loop itself — both the end-to-end serving optimization and the per-kernel work underneath it, since GEAK’s e2e_workflow recursively drives kernel_workflow to author and tune the individual hot kernels worth fixing. See Hyperloom optimization loop.

GEAK documentation#

For detailed documentation on GEAK, see GEAK on ROCm Docs.

Contents