GEAK documentation#
2026-07-22
1 min read time
GEAK (Generating Efficient AI-Centric Kernels) is a multi-agent GPU performance optimizer for
AMD Instinct™ MI GPUs (CDNA). It ships two deterministic Workflows: e2e_workflow raises the
end-to-end sglang or vLLM serving throughput of a whole model, and kernel_workflow optimizes a
single AMD GPU kernel (Triton, HIP, CK, FlyDSL). Control flow is deterministic JS; LLM agents are called
only for judgment.
The GEAK public repository is located at AMD-AGI/GEAK.
For information on contributing to the GEAK code base, see the CONTRIBUTING guide.
GEAK is released under the Apache License 2.0. For details, see the License page.