torch::autograd::validate_outputs needs to use guard oblivious test #119476
Description
🐛 Describe the bug
The is_same_shape call doesn't use size oblivious / expect true so it will choke if there are unbacked SymInts involved.
You can work around it with something like this:
diff --git a/fbcode/caffe2/torch/csrc/autograd/engine.cpp b/fbcode/caffe2/torch/csrc/autograd/engine.cpp
--- a/fbcode/caffe2/torch/csrc/autograd/engine.cpp
+++ b/fbcode/caffe2/torch/csrc/autograd/engine.cpp
@@ -847,14 +847,7 @@
continue;
}
- if (!metadata.is_same_shape(grad)) {
- if (metadata.is_expandable_to_shape(grad)) {
- grad = metadata.reduce_grad(grad);
- } else {
- const auto message = metadata.incompatible_shape_error_message(i, grad);
- TORCH_CHECK(false, format_error(message.str()));
- }
- }
+ grad = metadata.reduce_grad(grad);
bool input_is_complex =
isComplexType(c10::typeMetaToScalarType(metadata.options().dtype()));
Error looks something like:
It appears that you're trying to get a value out of symbolic int/float whose value is data-dependent (and thus we do not know the true value.) The expression we were trying to evaluate is Eq(i2 + i3, i4 + i5) (unhinted: Eq(i2 + i3, i4 + i5)). Scroll up to see where each of these data-dependent accesses originally occurred.
Note that it looks like maybe i4 and i5 are reallocations of i2 and i3, but these reallocations were actually done in user code (not in the framework) because they are repeatedly tolist()'ing the same tensor.
Versions
main
cc @albanD @zou3519 @gqchen @pearu @nikitaved @soulitzer @lezcano @Varal7 @msaroufim @bdhirsh @anijain2305