trunk/8368a088aeaec7fd68f9eef843876de15748c3be: Fix assume_constant_result tensor constants after graph breaks (#185373)
- PyTorch: 660 events in the last 90 days
- PyTorch: 651th Release in the last 90 days
- Previous: earlier the same day · trunk/2b1d72fc614c57d5ebe70ebaf22da2abaf11a3d4
What happened
invoke_and_store_as_constant is shared by assume_constant_result , builtin constant folding, and constant NNModule methods. When one of these paths returned a Tensor, Dynamo registered it as a graph attribute with a ConstantSource derived from the call-site name. Unlike the non-Tensor path, tensor registration did not install that name in the runtime globals. When bytecode reconstruction needed the value across a gr…
Summary assembled by rule from the sources below