April 2026 monthly summary focused on stabilizing the FP8 path in PyTorch by addressing Infinity handling for Float8_e4m3fn. Implemented a dedicated isinf() method to prevent NaN assertion crashes with FP8 inputs and to enforce the spec that Infinity is excluded in this format. This work reduces runtime crashes when nan_asserts are enabled in compiled models and enhances reliability of FP8 training/inference paths. Key outcomes include reduced error surface in FP8 workflows, enabling safer deployment and broader FP8 adoption across training and inference. The work aligns with the PyTorch FP8 design (e4m3fn) and contributes to overall model stability in production-like scenarios. Committed work and traceability: - Commit: 35012ea890770a8144504e6fbcd2ff0420e10ea6 - Pull Request: https://github.com/pytorch/pytorch/pull/160641 - PR resolves: #149002 and related nan_asserts runtime issue.