Automatically raise the serving throughput of a large language model running on sglang or vLLM.
Speed up one specific AMD GPU kernel written in Triton, HIP, or CK without manual profiling.
Get a written report explaining which optimizations were tried and why each one was accepted or rejected.
Batch-optimize many GPU kernels at once across multiple GPUs with automatic GPU locking.
| aakashvarma/geak | 0xallam/my-recipe | 0xhassaan/nn-from-scratch | |
|---|---|---|---|
| Stars | 0 | — | 0 |
| Language | Python | Python | Python |
| Last pushed | — | 2022-11-22 | — |
| Maintenance | — | Dormant | — |
| Setup difficulty | hard | moderate | moderate |
| Complexity | 5/5 | 2/5 | 4/5 |
| Audience | researcher | general | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires an AMD Instinct MI GPU, ROCm 6+, and a pre-built kernel toolchain such as PyTorch or Triton set up beforehand.
GEAK is a tool that uses AI agents to automatically speed up code that runs on AMD Instinct MI GPUs, the specialized chips used for running large AI models. It is built on top of Claude Code, Anthropic's coding assistant, and uses a system of scripted workflows to coordinate multiple AI agents working together on a performance problem, rather than relying on one agent to figure everything out unsupervised. The project ships two related tools. The main one, called e2e_workflow, looks at an entire AI model being served to real users through a serving system such as sglang or vLLM, and works to raise its overall throughput, meaning how many requests it can handle per second. It does this by first measuring where the model actually spends its GPU time, then trying the cheapest fixes first, such as adjusting configuration settings, before moving on to rewriting or generating faster low-level GPU code for the parts that matter most. The second tool, kernel_workflow, works on a single piece of GPU code at a time, called a kernel, and is what the first tool calls automatically whenever it decides a specific kernel needs to be rewritten for speed. Every change the system makes is checked before being accepted. It compares the before and after performance directly on the same hardware, confirms the output is still correct, and only keeps a change if it delivers a real, measurable improvement. Every run produces a written report describing what was tried, what worked, and what the final result was, so a person can review the reasoning afterward rather than trusting a black box. Getting started requires an AMD Instinct MI GPU, the ROCm software stack that lets you use it, and Python. You install GEAK with a single command, set up API access to an AI model, then launch Claude Code and simply describe in plain language what you want optimized, such as naming a model and its serving settings, or pointing at one specific piece of GPU code. The project reports that its kernel-level tool roughly doubled the speed gains achieved by an earlier version of the same idea, based on tests across a dozen sample kernels. This is aimed at machine learning engineers working on AMD hardware who want to speed up model serving without manually profiling and rewriting GPU code by hand.
GEAK uses coordinated Claude Code AI agents to automatically profile and speed up AI model serving and individual GPU kernels on AMD Instinct MI GPUs.
Mainly Python. The stack also includes Python, Claude Code, ROCm.
Setup difficulty is rated hard, with roughly 1h+ to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
double-check against the repo, no cap.