who-ate-my-flops

A small plugin for Claude Code and Codex to diagnose and optimize end-to-end performance in PyTorch training and inference workloads.

Blog · Installation · Quick Start · Roadmap

🔥 Optimizations developed with the plugin have been merged into FunASR, FastVideo, gsplat, Ultralytics, Unsloth, and SGLang, with measured speedups of up to 3.6× on the tested workloads.

See how it works in 1 minute

who-ate-my-flops-usage.mp4

What It Does

Provides a harness for end-to-end PyTorch job performance optimization, with:

Workload context and profiling analysis to guide the agent’s reasoning
Correctness checks to evaluate changes
Guidance for the agent to ask questions that uncover users’ goals and constraints.

Optimizations can range from configuration changes to Python code and GPU kernels. The plugin works on individual training or inference jobs with a repeatable launch command, e.g. jobs submitted through Slurm.

Installation

Ask your agent

Tell Claude Code or Codex:

Install the who-ate-my-flops plugin from
https://github.com/OpenPerfAgent/who-ate-my-flops
for this client. Follow the installation instructions in its README.

Manual installation

Claude Code

Run these commands in Claude Code:

/plugin marketplace add OpenPerfAgent/who-ate-my-flops
/plugin install who-ate-my-flops@openperfagent

Codex

Run these commands in your terminal:

codex plugin marketplace add OpenPerfAgent/who-ate-my-flops
codex plugin add who-ate-my-flops@openperfagent

Quick Start

Have your code repository, a command that launches the training or inference job, and access to an idle GPU environment ready. Start Claude Code or Codex in the repository you want to optimize.

Before running init, enable Auto mode in Claude Code or Approve for me in Codex so the agent can work with fewer interruptions.

1. Set up the workload

Run init and give the agent your launch command and GPU environment. It will ask about your optimization goal, constraints, and correctness requirements, and record them in contract.md.

Claude Code	Codex
`/who-ate-my-flops:init`	`$who-ate-my-flops:init`

2. Diagnose or optimize

Choose diagnose for one investigation and a measured fix, or optimize to let the agent work through multiple improvements.

	Claude Code	Codex
Diagnose	`/who-ate-my-flops:diagnose`	`$who-ate-my-flops:diagnose`
Optimize	`/who-ate-my-flops:optimize`	`$who-ate-my-flops:optimize`

3. Wait and review the results

Wait for the agent to finish, then start with latest-report.md to review the results. Reports and experiment records are in workspace-who-ate-my-flops/.

workspace-who-ate-my-flops/
├── latest-report.md   # Results, correctness checks, and reproduction commands
├── benchmark.csv      # Baseline and optimization measurements
├── contract.md        # Agreed goals and constraints
├── records/           # Process records and planning notes
├── tools/             # Helper scripts
├── runs/              # Run outputs
└── commits/           # Profiling traces and diagnoses by commit

Questions, bugs, or feedback? Open an issue.

Roadmap

Add skills for NVIDIA Nsight Systems.
Further improvement of the alignment with user intent.
Enable recursive delegation to kernel optimization agents when needed.

Citation

You can cite our blog post with:

@misc{zhao2026whoatemyflops,
  title        = {Beyond kernels: Letting agents optimize the whole {AI} job},
  author       = {Zhao, Hexu and Xi, Haocheng and Wang, Yichuan and Yin, Shaofeng and Lv, Zhaoyang and Feng, Haiwen and Li, Xiuyu and Panda, Aurojit and Li, Jinyang},
  year         = {2026},
  url          = {https://openperfagent.github.io/who-ate-my-flops/}
}

License

Licensed under the Apache License, Version 2.0.

Star growth

README