Handle MKLDNN to_dense() in FakeTensor and Inductor lowering so compiled MKLDNN inputs transition to strided dense tensors before arithmetic. The global aten.to_dense composite also covers the direct (non-method) densify path, so all reported "Cannot access data pointer" failures on MKL-DNN to_dense() under torch.compile are fixed.
Fixes #182402
Fixes #183773
Fixes #160873
Generated by my agent
Pull Request resolved: #183670
Approved by: https://github.com/aorenste
SOCIAL SHARE CARD GENERATOR