GitLab

* Adding _tie_weights() to prediction heads to support low_cpu_mem_usage=True

* Testing for the non-safe-tensors case, since the default is safe-tensors already

* Running fixup/fix-copies

* Adding accelerate annotations to tests