Remove duplicate check in distributions arg validation (#67741)
Summary:
Partial fix for https://github.com/pytorch/pytorch/issues/66800. (Duplicate of https://github.com/pytorch/pytorch/issues/67725 against pytorch/pytorch so as to trigger TorchBench)
https://github.com/pytorch/pytorch/issues/61056 added a more verbose error message for distributions failing argument validation. However, it did not replace the earlier error check as was originally intended and was flagged by xuzhao9 as being the potential cause of a perf regression in `test_eval[soft_actor_critic-cuda-eager]`.
xuzhao9: Is there a way for me to check if this resolves the perf issue you mentioned?
cc VitalyFedyunin ngimel
Note that existing tests already check for the error message and should verify that the removed lines are redundant.
RUN_TORCHBENCH: soft_actor_critic
Pull Request resolved: https://github.com/pytorch/pytorch/pull/67741
Reviewed By: neerajprad
Differential Revision: D32135675
Pulled By: xuzhao9
fbshipit-source-id: 37dfd3ff53b95017c763371979ab3a2c302a72b9
diff --git a/torch/distributions/distribution.py b/torch/distributions/distribution.py
index 38d8e38..b0f9d2e 100644
--- a/torch/distributions/distribution.py
+++ b/torch/distributions/distribution.py
@@ -59,8 +59,6 @@
f"to satisfy the constraint {repr(constraint)}, "
f"but found invalid values:\n{value}"
)
- if not constraint.check(getattr(self, param)).all():
- raise ValueError("The parameter {} has invalid values".format(param))
super(Distribution, self).__init__()
def expand(self, batch_shape, _instance=None):
