How to invoke symbolic shape propagation?

The faketensor colab Google Colab has an example of trying maximum batchsize, but it uses concrete shapes with trial and error.

How to create a faketensor with symbolic shape and pass it to a module to get the output symbolic shape?

The documentation Fake tensor — PyTorch 2.1 documentation mentioned ShapeEnv, but I can’t find a simple usage.

Might be relevant: @ezyang @albanD @Chillee

This sort of general questions are better suited for the general PyTorch forum https://discuss.pytorch.org/

At any rate, for this and similar questions, it’s always easy to find examples in the testing suite, that is test/. In this case, you can find plenty of examples in test/test_dynamic_shapes.py.

1 Like

@Lezcano Thanks for the pointer. It is indeed very helpful.