Reproduce “ModelBlaster: Closing the Performance Gap through Learning-Augmented Agents Across Models, Kernels and Hardware Backends” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shengjun Kris Dong, Dima Nikiforov, Sahil Modi, Sana Damani, Edward C Lin, Siva Kumar Sastry Hari, Christos Kozyrakis
ModelBlaster uses a persistent, profile-grounded memory to guide kernel optimization and whole-model graph choices across hardware backends. An agent records bottlenecks, attempted changes, outcomes, and constraints; candidates are compiled, checked for correctness, and measured before memory informs later attempts. On L40S KernelBench Level 2, it achieved 5.576× geometric-mean speedup over PyTorch eager, versus 2.177× for StitchCUDA, on this benchmark setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.