PyTorch optimizes Triton max/min reductions to reuse paired indexing

PyTorch's latest pull request (#184149) introduces a key optimization to its Triton backend, targeting max and min dimension reductions. Previously, reductions like `torch.max(dim)` and `torch.min(dim)` were handled with separate code paths, leading to duplication and potential maintenance overhead. Contributor jansel devised a method to internally convert these operations into arg-value reductions, which then reuse the existing paired indexed reduction logic. This change preserves the distinct semantics of `amax` and `amin` while consolidating implementation.

The PR, tagged as viable/strict/1780293447, was merged on June 1 and fixes issue #146643. By reducing code redundancy, the update improves the efficiency of PyTorch's Triton backend for GPU computations. For developers using PyTorch with Triton, this means fewer edge cases and a more unified codebase, potentially leading to faster compilation and easier debugging. The optimization is a behind-the-scenes improvement that strengthens PyTorch's support for advanced GPU kernels.