Attention.Config(
                channels_in=16,
                channels_out=16,
                num_heads=2,
                channels_head=8,
                num_heads_kv=2,
                bias=False,
                dropout=0.0,
                causal=False,
                rope=None,
                norm_qk=None,
                share_qk_norm=True,
                norm_out=None,
                split_qkv_projection=False,
                init_weight=<function priml.model.init.kaiming_uniform at 0xdefacedeface>,
                depth_index=(),
                attn_kernel=SdpaFused.Config()
        )
