Qwen3.Config(
        channels_in=8,
        channels_out=9,
        num_layers=1,
        proj_in=Embedding.Config(
            channels_in=9,
            channels_out=8,
            padding_idx=None,
            device=None,
            dtype=None,
            lookup_dtype=None,
            shard='vocab',
            depth_index=(),
            init_weight=functools.partial(torch.nn.init.normal_, std=0.02)
        ),
        block=[
        │   TransformerBlock.Config(
        │       channels_in=8,
        │       channels_out=8,
        │       attn=Attention.Config(
        │           channels_in=8,
        │           channels_out=8,
        │           num_heads=4,
        │           channels_head=6,
        │           num_heads_kv=2,
        │           bias=False,
        │           dropout=0.0,
        │           causal=True,
        │           rope=RoPE.Config(
        │               channels_head=6,
        │               reduction_mode='cat',
        │               frequencies=HuggingFaceFrequencies.Config(base=1000000.0),
        │               dtype=torch.float32
        │           ),
        │           rotate=None,
        │           norm_qk=RMSNorm.Config(
        │               channels_in=6,
        │               channels_out=6,
        │               eps=1e-06,
        │               elementwise_affine=True,
        │               device=None,
        │               dtype=None
        │           ),
        │           share_qk_norm=False,
        │           norm_out=None,
        │           split_qkv_projection=False,
        │           proj_qkv=None,
        │           proj_out=None,
        │           init_weight=functools.partial(torch.nn.init.normal_, std=0.02),
        │           depth_index=((0, 1),),
        │           attn_kernel=SdpaFused.Config()
        │       ),
        │       ffn=SwiGLU.Config(
        │           channels_in=8,
        │           channels_out=8,
        │           channels_hidden=7,
        │           expansion=2.6666666666666665,
        │           round_to=256,
        │           gate=True,
        │           bias=False,
        │           split_gate_projection=False,
        │           norm=None,
        │           act=priml.model.swiglu.silu,
        │           depth_index=((0, 1),),
        │           init_weight=functools.partial(torch.nn.init.normal_, std=0.02),
        │           init_weight_out=functools.partial(torch.nn.init.normal_, std=0.02),
        │           up_proj=None,
        │           middle=None,
        │           down_proj=None,
        │           shard='colwise'
        │       ),
        │       norm1=RMSNorm.Config(
        │           channels_in=8,
        │           channels_out=8,
        │           eps=1e-06,
        │           elementwise_affine=True,
        │           device=None,
        │           dtype=None
        │       ),
        │       norm2=RMSNorm.Config(
        │           channels_in=8,
        │           channels_out=8,
        │           eps=1e-06,
        │           elementwise_affine=True,
        │           device=None,
        │           dtype=None
        │       ),
        │       prenorm=True,
        │       checkpoint=False,
        │       depth_index=((0, 1),)
        │   )
        ],
        proj_out=Sequential.Config(
            channels_in=8,
            channels_out=9,
            elements=[
                RMSNorm.Config(
                    channels_in=8,
                    channels_out=8,
                    eps=1e-06,
                    elementwise_affine=True,
                    device=None,
                    dtype=None
                ),
                Linear.Config(
                    channels_in=8,
                    channels_out=9,
                    bias=False,
                    depth_index=(),
                    device=None,
                    dtype=None,
                    init_weight=functools.partial(torch.nn.init.normal_, std=0.02),
                    init_bias=torch.nn.init.zeros_,
                    linear_fn=<built-in function linear>,
                    shard='vocab'
                )
            ],
            repeat=1,
            depth_index=()
        )
    )
