tree 4b079d3899ad231ce66ece9200d16c00d9a1f786
parent 24d3ea61ab6dc4429851af258a2cdc8d3b52a9aa
author John Clow <jclow@fb.com> 1659975075 -0700
committer PyTorch MergeBot <pytorchmergebot@users.noreply.github.com> 1660057560 +0000

[SSA] Clear Shape Cache on Load Error (#82850)

If we do not clear when loading shape cache items errors out, it instead produces unhelpful errors on retry.
Every single subsequent test errors out with `method 'unary' already defined`.

See this example error in PR #80860:

```
ERROR [0.000s]: test_zerodim_cpu (jit.test_device_analysis.TestDeviceAnalysis) [In device: (device(type='cpu'), device(type='cpu')), expected: cpu,
 mul
 shapes: ((1, 2, 2), (2, 2)),
 devices: (device(type='cpu'), device(type='cpu'))]
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/var/lib/jenkins/workspace/test/jit/test_device_analysis.py", line 66, in assert_device_equal
    self.prop_device_on_graph(graph, in_devices, in_shapes)
  File "/var/lib/jenkins/workspace/test/jit/test_device_analysis.py", line 55, in prop_device_on_graph
    torch._C._jit_pass_propagate_shapes_on_graph(graph)
RuntimeError: expected eof but found ':' here:
aten::convolution : (Tensor, Tensor, Tensor?, int[], int[], int[], bool, int[], int) -> (Tensor)
                  ~ <--- HERE

======================================================================
ERROR [0.000s]: test_zerodim_cpu (jit.test_device_analysis.TestDeviceAnalysis) [In device: (device(type='cpu'), device(type='cpu')), expected: cpu,
 mul
 shapes: ((1, 2, 2), ()),
 devices: (device(type='cpu'), device(type='cpu'))]
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/var/lib/jenkins/workspace/test/jit/test_device_analysis.py", line 66, in assert_device_equal
    self.prop_device_on_graph(graph, in_devices, in_shapes)
  File "/var/lib/jenkins/workspace/test/jit/test_device_analysis.py", line 55, in prop_device_on_graph
    torch._C._jit_pass_propagate_shapes_on_graph(graph)
RuntimeError: method 'unary' already defined.
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82850
Approved by: https://github.com/davidberard98