Open amalbasaTT opened 1 week ago
@amalbasaTT For this bug do you mean that we are not supposed to allow different data types for output and input gradient tensors and we should exit with an error message ? Or do you mean different data types should be able to work together ?
@amalbasaTT For this bug do you mean that we are not supposed to allow different data types for output and input gradient tensors and we should exit with an error message ? Or do you mean different data types should be able to work together ?
I mean that this constraint should be in the documentation. And I think that different data types should be supported. I fixed the issue text so that the description of the problem isn't so vague, my bad on that.
ttnn.embedding_bw doesn't support output and input gradient tensors having different dtypes. This bug/constraint should be put in the documentation. . To Reproduce Steps to reproduce the behavior: Sweep test for embedding_bw is located in 'tests/sweep_framework/sweeps/eltwise/binary_backward/embedding_bw/embedding_bw.py'
https://github.com/tenstorrent/tt-metal/tree/main/tests/sweep_framework
Expected behavior For vectors where
output_dtype
andgrad_dtype
are different, test will fail with the message "Output and input gradient tensors must have the same dtype"