Reproduce “Accelerating the Mitigation of LLM Inference Nondeterminism Across GPU Architectures” 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
Liam Cooper, Shinnung Jeong, Hyeran Jeon, Jeffrey Young, Hyesoon Kim
ReproFuse makes linear-layer computations reproducible across tested GPUs by loading BF16 weights, widening them in registers, and using fixed-order FP32 fused multiply-add kernels. In Qwen3-4B GSM8K, cross-GPU divergence fell to 0.08%, from 70.81% with BF16 and 0.23% with LayerCast. The remaining case involved attention, so end-to-end inference is not guaranteed bitwise identical.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.