Closed Craigacp closed 7 months ago
This improves determinism and makes the gradients be computed correctly. As it's a control dependency on all the gradients to the consumers of the gradients it's not clear why this is necessary, but it's working.
Fix developed by @nfeybesse.
This improves determinism and makes the gradients be computed correctly. As it's a control dependency on all the gradients to the consumers of the gradients it's not clear why this is necessary, but it's working.
Fix developed by @nfeybesse.