Reproduce “PipeLink: Contract-Guided Linking of GPU Kernel Components” 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

Jize Guo, Xupeng Miao

PipeLink links separately compiled GPU-kernel components by checking contracts that declare resource requirements and promises, then placing resources and generating connections. Across 51 configurations on H200, B300, and RTX 5090, linked kernels ran within 0.2% of same-source, single-unit controls. Replacing attention in a Llama-1B kernel yielded 2.16× throughput at 4096 tokens, but was 5% slower at 16 tokens; linking does not speed components.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.