← Back to events
ActiveAIRelease

trunk/e320689d0698ae12305459e36b87ab5a0c64fccc: [FakeTensor] Support sparse compressed tensor ctors (#192091)

Photo: PyTorch Releases

What happened

Fake impls for aten.sparse_compressed_tensor overloads run the meta kernel and restore the output device, instead of failing with a device mismatch. _to_dense fake impl now covers all sparse layouts, not just COO. FunctionalTensor wrapping a sparse tensor answers C++ layout() queries (dispatch_layout + prim.layout metadata fn); previously this hard-aborted via TensorImpl::layout_impl. Real-tensor propagation and FX…

Summary assembled by rule from the sources below

Why it's spreading

Sources