PyTorch's Inductor fixes uniform decomposition for non-contiguous views

PyTorch's Inductor compiler—the AI graph compiler that accelerates model training and inference—has shipped a targeted fix for a subtle correctness bug. Pull request #191709 resolves issue #191193, where uniform decomposition failed on non-contiguous views. The patch changes how the compiler generates code for these operations, now computing the storage length first and then using an as_strided operation to reshape back to the intended view. This avoids invalid memory access patterns and ensures compiled kernels match the semantics of eager PyTorch execution.

The fix is significant because non-contiguous tensors are common in real-world workloads—think transposed embeddings, sliced attention heads, or sparse feature maps. Prior to this patch, Inductor could silently generate incorrect kernels for such tensors, leading to numerical mismatches or crashes. With the change, developers using torch.compile can trust their models to produce correct results across a wider range of tensor layouts. Approved by PyTorch core maintainer jansel, the PR is now merged into the trunk branch, and its impact will be felt by anyone building on PyTorch 2.x who relies on Inductor for graph compilation.