trunk/9454d796f69eee8ea179600f62681f98a4275e7a: [dynamic shapes] fix masked op DDEs (#173402)
A critical fix now allows 17 essential PyTorch functions to work with flexible AI model shapes.
Deep Dive
PyTorch developers have resolved a bug that prevented 17 'masked' operations from working with models that have flexible, unpredictable input shapes. These operations, including calculating sums, means, and softmax, are crucial for training and running modern AI. The fix handles these 'unbacked symbolic dimensions' by assuming shapes match during compilation and deferring final validation to runtime. This makes the framework more robust for complex, dynamic AI workloads.
Why It Matters
This fix removes a significant barrier for developers building AI models with adaptable and complex architectures.