Closed upsj closed 4 months ago
Noticed by @nbeams, the GCR initialization kernel uses an incorrect stride. This fixes it and adds a test for the behavior.
@yhmtsai exactly, the residual uses the default stride, while b can have an arbitrary stride
Error: PR already merged!
Noticed by @nbeams, the GCR initialization kernel uses an incorrect stride. This fixes it and adds a test for the behavior.