This PR is part of a series attempting to re-submit #134592 as smaller PRs.
In fx tests:
- Add and use a common raise_on_run_directly method for when a user runs a test file directly which should not be run this way. Print the file which the user should have run.
- Raise a RuntimeError on tests which have been disabled (not run)
- Remove any remaining uses of "unittest.main()""
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154715
Approved by: https://github.com/Skylion007
Summary:
Add a shape inference tool that helps to infer each node shape of a given graph module.
1. Given a fx graph, and an example of an input(don't need to be an accurate input that can be forward, but should have valid dims and data structures), `infer shape` creates an input of symbolic shape
2. Shape prop this symbolic input can catch runtime or value exceptions.
3. These errors are constraints for symbol values, and the constraint solver `infer symbolic values` helps us figure out specific values for each symbol.
4. Finally, we run the shape propagation based on input tensor to get tensor shapes for all nodes in the FX traced module.
Test Plan:
### 1. Test `infer symbol values`
Command:
```
buck2 test mode/opt //caffe2/test:fx_experimental -- test_infer_symbol_values
```
### 2. Test `infer shape`
Command:
```
buck2 test mode/opt //caffe2/test:fx_experimental -- test_infer_symbol_values
```
Inferred shape result like: P897560514
Differential Revision: D53593702
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120097
Approved by: https://github.com/yf225
