MinGRUBlock.Config(
        channels_hidden=4,
        proj_gates=Linear.Config(
            channels_in=4,
            channels_out=12,
            bias=False,
            depth_index=(),
            device=None,
            dtype=None,
            init_weight=priml.model.init.kaiming_uniform,
            init_bias=torch.nn.init.zeros_,
            linear_fn=<built-in function linear>,
            shard=None
        ),
        scan=TorchScan.Config(),
        dtype=None
    )
