What is GEAK?#
2026-07-17
2 min read time
GEAK (Generating Efficient AI-Centric Kernels) is a multi-agent GPU performance optimizer for AMD Instinct™ MI GPUs (CDNA; the on-box card is auto-detected). It ships two deterministic Workflows, driven by Claude Code:
Workflow |
Scope |
What it optimizes |
|---|---|---|
|
Whole-model serving |
End-to-end sglang or vLLM throughput of a full LLM |
|
Single kernel |
Latency or speedup of one AMD GPU kernel (Triton, HIP, CK, FlyDSL) |
e2e_workflow is the headline: it raises serving throughput by triaging hot kernels and pulling levers
cheapest-first, then recursively calls kernel_workflow to author or optimize the kernels worth
fixing. To speed up a single kernel, use kernel_workflow directly.
Core design: deterministic control plane, LLM judgment#
Control flow — the budget loop, parallel fan-out, verification, and stop conditions — is deterministic
JS in e2e_workflow.js / kernel_workflow.js. LLM agents are invoked only for judgment (analysis,
strategy, kernel authoring). This makes runs reliable and reproducible.
e2e_workflow (whole-model serving)#
A system layer that wraps — and recursively calls — the single-kernel layer. Specialized agents own each stage: an e2e Director (isolated environment + true baseline), a System Architect (Amdahl strategy), a Profiler, a Config Tuner, a Kernel Extractor, an Op Benchmarker, and an e2e Integrator (reversible overlay + throughput gate).
kernel_workflow (single kernel)#
A hierarchical single-kernel optimizer: Director → TechLead → specialist engineers (algorithm, memory, compute, host-runtime), multi-round and budget-controlled. Each patch is independently verified against an immutable correctness oracle before it is accepted.
Accuracy and measurement#
Profiling grounds every decision in measured data — nothing is optimized unless it is first shown to matter (Amdahl:
pct_gpu_time × achievable_speedup).A curated knowledge base (
perf_knowledge/) supplies operator × backend priors as reference only — the on-box benchmark and end-to-end gate are always the judge.
How a run works#
Preflight the environment (GPU arch, serving backend, model).
Profile a running server on your exact workload and rank hot kernels by Amdahl.
Pull levers cheapest-first — config/backend sweep → head GEMM/attention bake-off (aiter per-shape tune + a kernel authored using the recursive kernel layer) → editable-kernel milestone loop.
Overlay each accepted change back reversibly, gated on a measured warm-server throughput delta (interleaved A/B, output parity).
Report and validate — every run writes a complete
final_report.md, and the Director independently re-measures the result.