trunk/2b8b4ff5ba5ca5064a81c03ddabff95feb7bd3b1

The PyTorch open-source team has deployed a crucial fix for a memory leak that was hampering AI development on Apple's M1, M2, and M3 chips. The issue, identified in the Metal Performance Shaders (MPS) backend, was located in the `getStridedMPSNDArray` function. This function failed to properly release an NSArray object, causing system memory to be consumed and not freed during tensor operations. For developers training or inferring with models on Mac, this leak could lead to crashes, slowdowns, and unstable performance, undermining the promise of native, high-performance AI on Apple Silicon.

The fix, committed by contributor Kaiters56, simply adds an `autorelease` call to the problematic NSArray copy. This single-line change ensures the macOS memory management system correctly disposes of the temporary array, preventing the leak. This patch is a significant quality-of-life improvement for the growing ecosystem of ML engineers and researchers using Macs as primary development machines. It makes local experimentation with frameworks like Hugging Face Transformers or running open-source models like Llama 3 more reliable and efficient, closing a gap in PyTorch's otherwise robust cross-platform support.