trunk/87efdb80e6690233dafaf7a186c7e8a5fadf6c14: Allow dispatch of LayerNormBackwardKernel on all devices. (#174385)
A small code change unlocks AI model training on a wider range of devices.
Deep Dive
A recent update to the PyTorch machine learning framework now allows a core training function, LayerNormBackwardKernel, to run on all hardware devices. Previously, it was restricted to just CPUs and NVIDIA GPUs (CUDA). This change, approved by core developers, means AI researchers and engineers can train models using this function on other accelerators like AMD GPUs or specialized AI chips, increasing flexibility and potentially speeding up development on diverse hardware setups.
Why It Matters
This lowers barriers for AI development by letting teams use their existing hardware more effectively.