Skip to content
Hacker News · AI·· 20 h agoAI score52

Lean-based audit finds rounding bugs in softmax and other operators across PyTorch, JAX and TensorFlow

When Rounding Breaks Machine Learning

AI summary

A Lean-based audit of 162 TorchLean operator entries, compared against PyTorch, JAX, TensorFlow, NumPy, CuPy, ONNX Runtime and MLX across roughly 780 million outputs, turned up floating-point rounding bugs in widely used code including PyTorch, JAX, TensorFlow Probability, diffusers, PyTorch Geometric and xLSTM, several never reported before. In the Lean sweep, 57,547 outputs (96.8%) matched the exact result rounded once, while 1,329 matched TorchLean's specified sequence of rounded operations. The work builds on FloatLib, a Lean library for arbitrary-precision floating-point arithmetic whose roundoff tactic proves error bounds, and the author notes that proving a formula correct over the reals does not establish its floating-point accuracy.

Source:Hacker News · AI · robertj1.com