Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Currently, during forward, topkgate gate linear layer is converted to fp32. This is forbidden since the linear layer params are a view into deepspeed's parameter flat buffer. To fix it, use torch.nn.functional.linear with gate.weight.float(). Signed-off-by: Moshe Island <misland@habana.ai>
- Loading branch information