Agent plugin /who-ate-my-flops speeds PyTorch jobs up to 3.6x, PRs merged into six OSS repos
ruilong_li · x · 2026-10-01
The team behind /who-ate-my-flops built a Claude Code / Codex plugin that extends agent optimization beyond individual GPU kernels to end-to-end training and inference jobs, tackling data transfers, CPU work, and synchronization bottlenecks.
- Provides profiling tools, workload context, and correctness checks so agents iterate a full job through a profile-fix-verify-measure loop
- Performance PRs produced with the tool were merged into FunASR, FastVideo, gsplat, Ultralytics, Unsloth, and SGLang, with measured speedups up to 3.6x
- The blog from Impossible, UC Berkeley, and NYU authors shares experiments, profiles, and lessons learned; the plugin repo is open source
More from coding & agent
- LibraryDesignBench tests whether AI agents can design and effectively use code libraries — a1zhang · 2026-10-01
- HQ launches as shared AI context layer, claiming nearly 1000 businesses onboard — vasuman · 2026-10-01
- Ace launches as a digital employee with its own computer, email and logins — garrytan · 2026-10-01
- Anthropic's cost-optimize skill audits Claude API spend per completed task — udmrzn · 2026-10-01
- Dev rant: CI vs GitHub — the fierce fight over who slows you down — garrytan · 2026-10-01
- universal-modder turns Claude Code into a modder for almost any PC game — Aryvyo · 2026-10-01