#triton

torch.compile over-reads offset views in Triton kernels

PyTorch can silently return wrong results when compiled kernels skip writing masked output regions.