-
Notifications
You must be signed in to change notification settings - Fork 1.7k
feat: Allow discontiguous inputs to the group_rms_norm. #4982
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: main
Are you sure you want to change the base?
Conversation
When input tensors are outputs of a split operation, the stride size along the batch dimension mismatches the size of dimension. For example: a, b = c.split([size_a, size_b], dim=1) c.shape[1] = size_a + size_b a.stride = (c.shape[1], 1); a_normed.stride = (size_a, 1) b.stride = (c.shape[1], 1); b_normed.stride = (size_b, 1) This PR adds support for the above case. Note that last dimension of all inputs must be discontiguous. Signed-off-by: Simeng Liu <[email protected]>
/bot run --disable-fail-fast |
PR_Github #7860 [ run ] triggered by Bot |
PR_Github #7860 [ run ] completed with state |
Signed-off-by: kduan <[email protected]>
/bot run --disable-fail-fast |
PR_Github #8172 [ run ] triggered by Bot |
PR_Github #8172 [ run ] completed with state |
Both failures from the pipeline are known issues. We can proceed with skip. |
/bot run --disable-fail-fast |
PR_Github #8386 [ run ] triggered by Bot |
PR_Github #8386 [ run ] completed with state |
feat: Allow discontiguous inputs to the group_rms_norm.
Description
When input tensors are outputs of a split operation, the stride size along the batch dimension mismatches the size of dimension.
For example:
This PR adds support for the above case. Note that last dimension of all inputs must be discontiguous.
Test Coverage
tests/unittest/_torch/test_group_rms_norm.py::test_group_rms_norm_with_different_IO_strides