trunk/8c8414e5c03f21b5405acc2fd9115f4448dcd08a: revert https://github.com/pytorch/pytorch/pull/172340 (#179151)

In a recent commit to the PyTorch main branch (trunk), Meta's AI engineering team performed a strategic rollback of a GPU performance optimization. The reverted change, identified as commit 8c8414e, had enabled "Lt bias fusions" to work with NVIDIA's cuBLAS library as the default computational backend. This type of fusion is a compiler technique that combines multiple GPU operations into a single, more efficient kernel, which is critical for speeding up deep learning training and inference on NVIDIA hardware.

The decision to revert was not taken lightly. The optimization was linked to a mysterious performance regression discovered in a separate pull request (#177703). The core issue is that while the fusion expanded coverage for many input types, it inadvertently "breaks things" in ways the team does "not quite know yet," as noted by engineer nikitaved. This highlights the delicate balance in high-performance computing: aggressive optimizations can have unpredictable side-effects across the vast matrix of possible model architectures and input sizes.

This incident is a classic example of the complex, iterative engineering required to maintain a foundational framework like PyTorch, which is used by millions of developers. The temporary rollback ensures stability for users while the root cause is diagnosed. The team indicated the change will become obsolete once a more comprehensive fix (PR #170571) is safely merged, demonstrating their methodical approach to deploying performance upgrades.