AION
Open-source releaseTraining & Scaling · Efficiency & Inference1 source · Sep 30, 2026

pytorch/pytorch v2.14.1: PyTorch 2.14.1 Release

This release is meant to fix the following regressions and silent correctness issues:

NVIDIAPyTorch
← Back to brief

Key points

  • cuBLAS: cublasLtMatmul() could ignore tensor-wide scaling for NVFP4 matrix multiplications (introduced in CUDA 13.2 Update 1)
  • Compiler: failed thread reconvergence could leave stale or corrupted register values in kernels with nested thread divergence (present since CUDA 12.8)

Sources (1)

  • [1]pytorch/pytorch v2.14.1: PyTorch 2.14.1 Release
    GitHub: pytorch/pytorch · Sep 30, 07:14 PM
    This release is meant to fix the following regressions and silent correctness issues:
    - cuBLAS: `cublasLtMatmul()` could ignore tensor-wide scaling for NVFP4 matrix multiplications (introduced in CUDA 13.2 Update 1)

Extractive summary: sentences quoted from the sources.