Added optional kind parameter to specify the model weight precision to most pipeline configurations. If not provided, will default to full precision on CPU, or the serialized weights precision otherwise.
Fixed issue with GPT-J that was incorrectly tracking the gradients for the attention bias
kindparameter to specify the model weight precision to most pipeline configurations. If not provided, will default to full precision on CPU, or the serialized weights precision otherwise.torch2.1 (viatch0.14.0).Fixes https://github.com/guillaume-be/rust-bert/issues/434