MultiHeadLatentAttention.Config(
        channels_in=32,
        channels_out=32,
        num_heads=4,
        channels_qk_nope_head=8,
        channels_qk_rope_head=4,
        channels_v_head=16,
        q_lora_rank=None,
        kv_lora_rank=12,
        bias=False,
        dropout=0.0,
        causal=True,
        softmax_scale=None,
        rope=None,
        attn_kernel=LatentAttention.Config(absorb=True, attn_kernel=SdpaNaive.Config()),
        norm_q_lora=RMSNorm.Config(
            channels_in=-1,
            channels_out=-1,
            eps=1e-06,
            elementwise_affine=True,
            device=None,
            dtype=None
        ),
        proj_q=Linear.Config(
            channels_in=32,
            channels_out=48,
            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
        ),
        proj_q_a=Linear.Config(
            channels_in=-1,
            channels_out=-1,
            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
        ),
        proj_q_b=Linear.Config(
            channels_in=-1,
            channels_out=-1,
            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
        ),
        proj_kv_a=Linear.Config(
            channels_in=32,
            channels_out=16,
            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
        ),
        proj_kv_b=Linear.Config(
            channels_in=12,
            channels_out=96,
            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
        ),
        proj_out=Linear.Config(
            channels_in=64,
            channels_out=32,
            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
        ),
        norm_kv_lora=RMSNorm.Config(
            channels_in=12,
            channels_out=12,
            eps=1e-06,
            elementwise_affine=True,
            device=None,
            dtype=None
        ),
        init_weight=priml.model.init.kaiming_uniform,
        depth_index=(),
        shard=None
    )
