program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.25.1"}})]
{
    func depth_in<ios18>(tensor<fp16, [1, 2048]> h) {
            tensor<fp16, [12288, 2048]> dd_depth_linear_weight_to_fp16 = const()[name = string("dd_depth_linear_weight_to_fp16"), val = tensor<fp16, [12288, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<fp16, [12288]> dd_depth_linear_bias_to_fp16 = const()[name = string("dd_depth_linear_bias_to_fp16"), val = tensor<fp16, [12288]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50331776)))];
            tensor<fp16, [1, 12288]> linear_0_cast_fp16 = linear(bias = dd_depth_linear_bias_to_fp16, weight = dd_depth_linear_weight_to_fp16, x = h)[name = string("linear_0_cast_fp16")];
            tensor<int32, [3]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [3]>([1, 12, 1024])];
            tensor<fp16, [1, 12, 1024]> din = reshape(shape = concat_0, x = linear_0_cast_fp16)[name = string("op_11_cast_fp16")];
        } -> (din);
    func depth_pass<ios18>(tensor<fp16, [1, 12, 1024]> slots) {
            bool var_26 = const()[name = string("op_26"), val = bool(true)];
            int32 var_27 = const()[name = string("op_27"), val = int32(-1)];
            tensor<fp16, [1, 12, 1024]> var_46_cast_fp16 = abs(x = slots)[name = string("op_46_cast_fp16")];
            tensor<int32, [1]> var_48_axes_0 = const()[name = string("op_48_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 1]> var_48_cast_fp16 = reduce_max(axes = var_48_axes_0, keep_dims = var_26, x = var_46_cast_fp16)[name = string("op_48_cast_fp16")];
            fp16 var_25_to_fp16 = const()[name = string("op_25_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 12, 1]> s_1_cast_fp16 = maximum(x = var_48_cast_fp16, y = var_25_to_fp16)[name = string("s_1_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> y_1_cast_fp16 = real_div(x = slots, y = s_1_cast_fp16)[name = string("y_1_cast_fp16")];
            fp16 var_24_promoted_to_fp16 = const()[name = string("op_24_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 1024]> var_51_cast_fp16 = pow(x = y_1_cast_fp16, y = var_24_promoted_to_fp16)[name = string("op_51_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("op_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_53_keep_dims_0 = const()[name = string("op_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 1]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_51_cast_fp16)[name = string("op_53_cast_fp16")];
            fp32 var_54_epsilon_0 = const()[name = string("op_54_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 12, 1]> var_54_cast_fp16 = inverse(epsilon = var_54_epsilon_0, x = s_1_cast_fp16)[name = string("op_54_cast_fp16")];
            fp16 var_55_to_fp16 = const()[name = string("op_55_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 1]> var_56_cast_fp16 = mul(x = var_54_cast_fp16, y = var_55_to_fp16)[name = string("op_56_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_57_cast_fp16 = real_div(x = var_56_cast_fp16, y = s_1_cast_fp16)[name = string("op_57_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_58_cast_fp16 = add(x = var_53_cast_fp16, y = var_57_cast_fp16)[name = string("op_58_cast_fp16")];
            fp32 var_59_epsilon_0 = const()[name = string("op_59_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 1]> var_59_cast_fp16 = rsqrt(epsilon = var_59_epsilon_0, x = var_58_cast_fp16)[name = string("op_59_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_60_cast_fp16 = mul(x = y_1_cast_fp16, y = var_59_cast_fp16)[name = string("op_60_cast_fp16")];
            tensor<fp16, [1024]> dd_trunk_layers_0_operator_norm_weight_to_fp16 = const()[name = string("dd_trunk_layers_0_operator_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50356416)))];
            tensor<fp16, [1, 12, 1024]> input_1_cast_fp16 = mul(x = dd_trunk_layers_0_operator_norm_weight_to_fp16, y = var_60_cast_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1024, 1024]> dd_trunk_layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50358528)))];
            tensor<fp16, [1024]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52455744)))];
            tensor<fp16, [1, 12, 1024]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = dd_trunk_layers_0_self_attn_q_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [4]>([1, 12, 32, 32])];
            tensor<fp16, [1, 12, 32, 32]> x_1_cast_fp16 = reshape(shape = concat_0, x = linear_0_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> var_74_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_74_cast_fp16")];
            tensor<int32, [1]> var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 32, 1]> var_76_cast_fp16 = reduce_max(axes = var_76_axes_0, keep_dims = var_26, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")];
            tensor<fp16, [1, 12, 32, 1]> s_3_cast_fp16 = maximum(x = var_76_cast_fp16, y = var_25_to_fp16)[name = string("s_3_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> y_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = s_3_cast_fp16)[name = string("y_3_cast_fp16")];
            fp16 var_24_promoted_1_to_fp16 = const()[name = string("op_24_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 32, 32]> var_79_cast_fp16 = pow(x = y_3_cast_fp16, y = var_24_promoted_1_to_fp16)[name = string("op_79_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("op_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_81_keep_dims_0 = const()[name = string("op_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 32, 1]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_79_cast_fp16)[name = string("op_81_cast_fp16")];
            fp32 var_82_epsilon_0 = const()[name = string("op_82_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 12, 32, 1]> var_82_cast_fp16 = inverse(epsilon = var_82_epsilon_0, x = s_3_cast_fp16)[name = string("op_82_cast_fp16")];
            fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 32, 1]> var_84_cast_fp16 = mul(x = var_82_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")];
            tensor<fp16, [1, 12, 32, 1]> var_85_cast_fp16 = real_div(x = var_84_cast_fp16, y = s_3_cast_fp16)[name = string("op_85_cast_fp16")];
            tensor<fp16, [1, 12, 32, 1]> var_86_cast_fp16 = add(x = var_81_cast_fp16, y = var_85_cast_fp16)[name = string("op_86_cast_fp16")];
            fp32 var_87_epsilon_0 = const()[name = string("op_87_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 32, 1]> var_87_cast_fp16 = rsqrt(epsilon = var_87_epsilon_0, x = var_86_cast_fp16)[name = string("op_87_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> var_88_cast_fp16 = mul(x = y_3_cast_fp16, y = var_87_cast_fp16)[name = string("op_88_cast_fp16")];
            tensor<fp16, [32]> dd_trunk_layers_0_self_attn_q_norm_weight_to_fp16 = const()[name = string("dd_trunk_layers_0_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52457856)))];
            tensor<fp16, [1, 12, 32, 32]> var_89_cast_fp16 = mul(x = dd_trunk_layers_0_self_attn_q_norm_weight_to_fp16, y = var_88_cast_fp16)[name = string("op_89_cast_fp16")];
            tensor<fp16, [256, 1024]> dd_trunk_layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52457984)))];
            tensor<fp16, [256]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52982336)))];
            tensor<fp16, [1, 12, 256]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = dd_trunk_layers_0_self_attn_k_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> concat_1 = const()[name = string("concat_1"), val = tensor<int32, [4]>([1, 12, 8, 32])];
            tensor<fp16, [1, 12, 8, 32]> x_3_cast_fp16 = reshape(shape = concat_1, x = linear_1_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 12, 8, 32]> var_96_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_96_cast_fp16")];
            tensor<int32, [1]> var_98_axes_0 = const()[name = string("op_98_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 8, 1]> var_98_cast_fp16 = reduce_max(axes = var_98_axes_0, keep_dims = var_26, x = var_96_cast_fp16)[name = string("op_98_cast_fp16")];
            tensor<fp16, [1, 12, 8, 1]> s_5_cast_fp16 = maximum(x = var_98_cast_fp16, y = var_25_to_fp16)[name = string("s_5_cast_fp16")];
            tensor<fp16, [1, 12, 8, 32]> y_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = s_5_cast_fp16)[name = string("y_5_cast_fp16")];
            fp16 var_24_promoted_2_to_fp16 = const()[name = string("op_24_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 8, 32]> var_101_cast_fp16 = pow(x = y_5_cast_fp16, y = var_24_promoted_2_to_fp16)[name = string("op_101_cast_fp16")];
            tensor<int32, [1]> var_103_axes_0 = const()[name = string("op_103_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_103_keep_dims_0 = const()[name = string("op_103_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 8, 1]> var_103_cast_fp16 = reduce_mean(axes = var_103_axes_0, keep_dims = var_103_keep_dims_0, x = var_101_cast_fp16)[name = string("op_103_cast_fp16")];
            fp32 var_104_epsilon_0 = const()[name = string("op_104_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 12, 8, 1]> var_104_cast_fp16 = inverse(epsilon = var_104_epsilon_0, x = s_5_cast_fp16)[name = string("op_104_cast_fp16")];
            fp16 var_105_to_fp16 = const()[name = string("op_105_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 8, 1]> var_106_cast_fp16 = mul(x = var_104_cast_fp16, y = var_105_to_fp16)[name = string("op_106_cast_fp16")];
            tensor<fp16, [1, 12, 8, 1]> var_107_cast_fp16 = real_div(x = var_106_cast_fp16, y = s_5_cast_fp16)[name = string("op_107_cast_fp16")];
            tensor<fp16, [1, 12, 8, 1]> var_108_cast_fp16 = add(x = var_103_cast_fp16, y = var_107_cast_fp16)[name = string("op_108_cast_fp16")];
            fp32 var_109_epsilon_0 = const()[name = string("op_109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 8, 1]> var_109_cast_fp16 = rsqrt(epsilon = var_109_epsilon_0, x = var_108_cast_fp16)[name = string("op_109_cast_fp16")];
            tensor<fp16, [1, 12, 8, 32]> var_110_cast_fp16 = mul(x = y_5_cast_fp16, y = var_109_cast_fp16)[name = string("op_110_cast_fp16")];
            tensor<fp16, [32]> dd_trunk_layers_0_self_attn_k_norm_weight_to_fp16 = const()[name = string("dd_trunk_layers_0_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52982912)))];
            tensor<fp16, [1, 12, 8, 32]> var_111_cast_fp16 = mul(x = dd_trunk_layers_0_self_attn_k_norm_weight_to_fp16, y = var_110_cast_fp16)[name = string("op_111_cast_fp16")];
            tensor<int32, [4]> x_7_perm_0 = const()[name = string("x_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 1024]> dd_trunk_layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52983040)))];
            tensor<fp16, [1, 12, 256]> linear_2_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = dd_trunk_layers_0_self_attn_v_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> concat_2 = const()[name = string("concat_2"), val = tensor<int32, [4]>([1, 12, 8, 32])];
            tensor<fp16, [1, 12, 8, 32]> var_116_cast_fp16 = reshape(shape = concat_2, x = linear_2_cast_fp16)[name = string("op_116_cast_fp16")];
            tensor<int32, [4]> v_1_perm_0 = const()[name = string("v_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 32]> const_8_to_fp16 = const()[name = string("const_8_to_fp16"), val = tensor<fp16, [1, 12, 1, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53507392)))];
            tensor<fp16, [1, 12, 32, 32]> var_118_cast_fp16 = mul(x = var_89_cast_fp16, y = const_8_to_fp16)[name = string("op_118_cast_fp16")];
            tensor<int32, [2]> var_119_split_sizes_0 = const()[name = string("op_119_split_sizes_0"), val = tensor<int32, [2]>([16, 16])];
            int32 const_9 = const()[name = string("const_9"), val = int32(3)];
            tensor<fp16, [1, 12, 32, 16]> var_119_cast_fp16_0, tensor<fp16, [1, 12, 32, 16]> var_119_cast_fp16_1 = split(axis = const_9, split_sizes = var_119_split_sizes_0, x = var_89_cast_fp16)[name = string("op_119_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, 32, 16]> var_121_cast_fp16 = mul(x = var_119_cast_fp16_1, y = const_2_promoted_to_fp16)[name = string("op_121_cast_fp16")];
            bool var_123_interleave_0 = const()[name = string("op_123_interleave_0"), val = bool(false)];
            int32 const_10 = const()[name = string("const_10"), val = int32(3)];
            tensor<fp16, [1, 12, 32, 32]> var_123_cast_fp16 = concat(axis = const_10, interleave = var_123_interleave_0, values = (var_121_cast_fp16, var_119_cast_fp16_0))[name = string("op_123_cast_fp16")];
            tensor<fp16, [1, 12, 1, 32]> const_11_to_fp16 = const()[name = string("const_11_to_fp16"), val = tensor<fp16, [1, 12, 1, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53508224)))];
            tensor<fp16, [1, 12, 32, 32]> var_124_cast_fp16 = mul(x = var_123_cast_fp16, y = const_11_to_fp16)[name = string("op_124_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> q_1_cast_fp16 = add(x = var_118_cast_fp16, y = var_124_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [1, 1, 12, 32]> const_0_to_fp16 = const()[name = string("const_0_to_fp16"), val = tensor<fp16, [1, 1, 12, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53509056)))];
            tensor<fp16, [1, 8, 12, 32]> x_7_cast_fp16 = transpose(perm = x_7_perm_0, x = var_111_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 8, 12, 32]> var_126_cast_fp16 = mul(x = x_7_cast_fp16, y = const_0_to_fp16)[name = string("op_126_cast_fp16")];
            tensor<int32, [2]> var_127_split_sizes_0 = const()[name = string("op_127_split_sizes_0"), val = tensor<int32, [2]>([16, 16])];
            int32 var_127_axis_0 = const()[name = string("op_127_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 12, 16]> var_127_cast_fp16_0, tensor<fp16, [1, 8, 12, 16]> var_127_cast_fp16_1 = split(axis = var_127_axis_0, split_sizes = var_127_split_sizes_0, x = x_7_cast_fp16)[name = string("op_127_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 12, 16]> var_129_cast_fp16 = mul(x = var_127_cast_fp16_1, y = const_3_promoted_to_fp16)[name = string("op_129_cast_fp16")];
            bool var_131_interleave_0 = const()[name = string("op_131_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 12, 32]> var_131_cast_fp16 = concat(axis = var_27, interleave = var_131_interleave_0, values = (var_129_cast_fp16, var_127_cast_fp16_0))[name = string("op_131_cast_fp16")];
            tensor<fp16, [1, 1, 12, 32]> const_1_to_fp16 = const()[name = string("const_1_to_fp16"), val = tensor<fp16, [1, 1, 12, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53509888)))];
            tensor<fp16, [1, 8, 12, 32]> var_132_cast_fp16 = mul(x = var_131_cast_fp16, y = const_1_to_fp16)[name = string("op_132_cast_fp16")];
            tensor<fp16, [1, 8, 12, 32]> k_1_cast_fp16 = add(x = var_126_cast_fp16, y = var_132_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<int32, [1]> var_134_axes_0 = const()[name = string("op_134_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 12, 32]> var_134_cast_fp16 = expand_dims(axes = var_134_axes_0, x = k_1_cast_fp16)[name = string("op_134_cast_fp16")];
            tensor<int32, [5]> var_136_reps_0 = const()[name = string("op_136_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 8, 4, 12, 32]> var_136_cast_fp16 = tile(reps = var_136_reps_0, x = var_134_cast_fp16)[name = string("op_136_cast_fp16")];
            tensor<int32, [4]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [4]>([1, 32, 12, 32])];
            tensor<fp16, [1, 32, 12, 32]> k_3_cast_fp16 = reshape(shape = concat_4, x = var_136_cast_fp16)[name = string("k_3_cast_fp16")];
            tensor<int32, [1]> var_138_axes_0 = const()[name = string("op_138_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 12, 32]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = var_116_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 8, 1, 12, 32]> var_138_cast_fp16 = expand_dims(axes = var_138_axes_0, x = v_1_cast_fp16)[name = string("op_138_cast_fp16")];
            tensor<int32, [5]> var_140_reps_0 = const()[name = string("op_140_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 8, 4, 12, 32]> var_140_cast_fp16 = tile(reps = var_140_reps_0, x = var_138_cast_fp16)[name = string("op_140_cast_fp16")];
            tensor<int32, [4]> concat_6 = const()[name = string("concat_6"), val = tensor<int32, [4]>([1, 32, 12, 32])];
            tensor<fp16, [1, 32, 12, 32]> v_3_cast_fp16 = reshape(shape = concat_6, x = var_140_cast_fp16)[name = string("v_3_cast_fp16")];
            tensor<int32, [4]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_143_transpose_x_1 = const()[name = string("op_143_transpose_x_1"), val = bool(false)];
            bool var_143_transpose_y_1 = const()[name = string("op_143_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 32, 12, 32]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = q_1_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 32, 12, 12]> var_143_cast_fp16 = matmul(transpose_x = var_143_transpose_x_1, transpose_y = var_143_transpose_y_1, x = transpose_0_cast_fp16, y = k_3_cast_fp16)[name = string("op_143_cast_fp16")];
            fp16 var_144_to_fp16 = const()[name = string("op_144_to_fp16"), val = fp16(0x1.6ap-3)];
            tensor<fp16, [1, 32, 12, 12]> x_9_cast_fp16 = mul(x = var_143_cast_fp16, y = var_144_to_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [1]> var_147_axes_0 = const()[name = string("op_147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 12, 1]> var_147_cast_fp16 = reduce_max(axes = var_147_axes_0, keep_dims = var_26, x = x_9_cast_fp16)[name = string("op_147_cast_fp16")];
            tensor<fp16, [1, 32, 12, 12]> x_11_cast_fp16 = sub(x = x_9_cast_fp16, y = var_147_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 32, 12, 12]> x_13_cast_fp16 = exp(x = x_11_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<int32, [1]> var_151_axes_0 = const()[name = string("op_151_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_151_keep_dims_0 = const()[name = string("op_151_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 12, 1]> var_151_cast_fp16 = reduce_sum(axes = var_151_axes_0, keep_dims = var_151_keep_dims_0, x = x_13_cast_fp16)[name = string("op_151_cast_fp16")];
            tensor<fp16, [1, 32, 12, 12]> var_152_cast_fp16 = real_div(x = x_13_cast_fp16, y = var_151_cast_fp16)[name = string("op_152_cast_fp16")];
            bool o_1_transpose_x_0 = const()[name = string("o_1_transpose_x_0"), val = bool(false)];
            bool o_1_transpose_y_0 = const()[name = string("o_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 32, 12, 32]> o_1_cast_fp16 = matmul(transpose_x = o_1_transpose_x_0, transpose_y = o_1_transpose_y_0, x = var_152_cast_fp16, y = v_3_cast_fp16)[name = string("o_1_cast_fp16")];
            tensor<int32, [4]> o_3_perm_0 = const()[name = string("o_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [32, 1024]> dd_trunk_layers_0_self_attn_gate_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_0_self_attn_gate_proj_weight_to_fp16"), val = tensor<fp16, [32, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53510720)))];
            tensor<fp16, [32]> linear_3_bias_0_to_fp16 = const()[name = string("linear_3_bias_0_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53576320)))];
            tensor<fp16, [1, 12, 32]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = dd_trunk_layers_0_self_attn_gate_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 12, 32]> var_158_cast_fp16 = sigmoid(x = linear_3_cast_fp16)[name = string("op_158_cast_fp16")];
            tensor<int32, [1]> var_159_axes_0 = const()[name = string("op_159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 32, 1]> var_159_cast_fp16 = expand_dims(axes = var_159_axes_0, x = var_158_cast_fp16)[name = string("op_159_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> o_3_cast_fp16 = transpose(perm = o_3_perm_0, x = o_1_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 12, 32, 32]> o_5_cast_fp16 = mul(x = o_3_cast_fp16, y = var_159_cast_fp16)[name = string("o_5_cast_fp16")];
            tensor<int32, [3]> concat_7 = const()[name = string("concat_7"), val = tensor<int32, [3]>([1, 12, 1024])];
            tensor<fp16, [1, 12, 1024]> input_3_cast_fp16 = reshape(shape = concat_7, x = o_5_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1024, 1024]> dd_trunk_layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53576448)))];
            tensor<fp16, [1, 12, 1024]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = dd_trunk_layers_0_self_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> x_17_cast_fp16 = add(x = slots, y = linear_4_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_166_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_166_cast_fp16")];
            tensor<int32, [1]> var_168_axes_0 = const()[name = string("op_168_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 1]> var_168_cast_fp16 = reduce_max(axes = var_168_axes_0, keep_dims = var_26, x = var_166_cast_fp16)[name = string("op_168_cast_fp16")];
            tensor<fp16, [1, 12, 1]> s_7_cast_fp16 = maximum(x = var_168_cast_fp16, y = var_25_to_fp16)[name = string("s_7_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> y_7_cast_fp16 = real_div(x = x_17_cast_fp16, y = s_7_cast_fp16)[name = string("y_7_cast_fp16")];
            fp16 var_24_promoted_3_to_fp16 = const()[name = string("op_24_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 1024]> var_171_cast_fp16 = pow(x = y_7_cast_fp16, y = var_24_promoted_3_to_fp16)[name = string("op_171_cast_fp16")];
            tensor<int32, [1]> var_173_axes_0 = const()[name = string("op_173_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_173_keep_dims_0 = const()[name = string("op_173_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 1]> var_173_cast_fp16 = reduce_mean(axes = var_173_axes_0, keep_dims = var_173_keep_dims_0, x = var_171_cast_fp16)[name = string("op_173_cast_fp16")];
            fp32 var_174_epsilon_0 = const()[name = string("op_174_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 12, 1]> var_174_cast_fp16 = inverse(epsilon = var_174_epsilon_0, x = s_7_cast_fp16)[name = string("op_174_cast_fp16")];
            fp16 var_175_to_fp16 = const()[name = string("op_175_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 1]> var_176_cast_fp16 = mul(x = var_174_cast_fp16, y = var_175_to_fp16)[name = string("op_176_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_177_cast_fp16 = real_div(x = var_176_cast_fp16, y = s_7_cast_fp16)[name = string("op_177_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_178_cast_fp16 = add(x = var_173_cast_fp16, y = var_177_cast_fp16)[name = string("op_178_cast_fp16")];
            fp32 var_179_epsilon_0 = const()[name = string("op_179_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 1]> var_179_cast_fp16 = rsqrt(epsilon = var_179_epsilon_0, x = var_178_cast_fp16)[name = string("op_179_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_180_cast_fp16 = mul(x = y_7_cast_fp16, y = var_179_cast_fp16)[name = string("op_180_cast_fp16")];
            tensor<fp16, [1024]> dd_trunk_layers_0_ffn_norm_weight_to_fp16 = const()[name = string("dd_trunk_layers_0_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55673664)))];
            tensor<fp16, [1, 12, 1024]> input_5_cast_fp16 = mul(x = dd_trunk_layers_0_ffn_norm_weight_to_fp16, y = var_180_cast_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [2816, 1024]> dd_trunk_layers_0_feed_forward_w1_weight_to_fp16 = const()[name = string("dd_trunk_layers_0_feed_forward_w1_weight_to_fp16"), val = tensor<fp16, [2816, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55675776)))];
            tensor<fp16, [2816]> linear_5_bias_0_to_fp16 = const()[name = string("linear_5_bias_0_to_fp16"), val = tensor<fp16, [2816]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61443008)))];
            tensor<fp16, [1, 12, 2816]> linear_5_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = dd_trunk_layers_0_feed_forward_w1_weight_to_fp16, x = input_5_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [2816, 1024]> dd_trunk_layers_0_feed_forward_w3_weight_to_fp16 = const()[name = string("dd_trunk_layers_0_feed_forward_w3_weight_to_fp16"), val = tensor<fp16, [2816, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61448704)))];
            tensor<fp16, [1, 12, 2816]> linear_6_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = dd_trunk_layers_0_feed_forward_w3_weight_to_fp16, x = input_5_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 12, 2816]> var_189_cast_fp16 = silu(x = linear_5_cast_fp16)[name = string("op_189_cast_fp16")];
            tensor<fp16, [1, 12, 2816]> input_9_cast_fp16 = mul(x = var_189_cast_fp16, y = linear_6_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1024, 2816]> dd_trunk_layers_0_feed_forward_w2_weight_to_fp16 = const()[name = string("dd_trunk_layers_0_feed_forward_w2_weight_to_fp16"), val = tensor<fp16, [1024, 2816]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67215936)))];
            tensor<fp16, [1, 12, 1024]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = dd_trunk_layers_0_feed_forward_w2_weight_to_fp16, x = input_9_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> x_19_cast_fp16 = add(x = x_17_cast_fp16, y = linear_7_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_199_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_199_cast_fp16")];
            tensor<int32, [1]> var_201_axes_0 = const()[name = string("op_201_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 1]> var_201_cast_fp16 = reduce_max(axes = var_201_axes_0, keep_dims = var_26, x = var_199_cast_fp16)[name = string("op_201_cast_fp16")];
            tensor<fp16, [1, 12, 1]> s_9_cast_fp16 = maximum(x = var_201_cast_fp16, y = var_25_to_fp16)[name = string("s_9_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> y_9_cast_fp16 = real_div(x = x_19_cast_fp16, y = s_9_cast_fp16)[name = string("y_9_cast_fp16")];
            fp16 var_24_promoted_4_to_fp16 = const()[name = string("op_24_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 1024]> var_204_cast_fp16 = pow(x = y_9_cast_fp16, y = var_24_promoted_4_to_fp16)[name = string("op_204_cast_fp16")];
            tensor<int32, [1]> var_206_axes_0 = const()[name = string("op_206_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_206_keep_dims_0 = const()[name = string("op_206_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 1]> var_206_cast_fp16 = reduce_mean(axes = var_206_axes_0, keep_dims = var_206_keep_dims_0, x = var_204_cast_fp16)[name = string("op_206_cast_fp16")];
            fp32 var_207_epsilon_0 = const()[name = string("op_207_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 12, 1]> var_207_cast_fp16 = inverse(epsilon = var_207_epsilon_0, x = s_9_cast_fp16)[name = string("op_207_cast_fp16")];
            fp16 var_208_to_fp16 = const()[name = string("op_208_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 1]> var_209_cast_fp16 = mul(x = var_207_cast_fp16, y = var_208_to_fp16)[name = string("op_209_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_210_cast_fp16 = real_div(x = var_209_cast_fp16, y = s_9_cast_fp16)[name = string("op_210_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_211_cast_fp16 = add(x = var_206_cast_fp16, y = var_210_cast_fp16)[name = string("op_211_cast_fp16")];
            fp32 var_212_epsilon_0 = const()[name = string("op_212_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 1]> var_212_cast_fp16 = rsqrt(epsilon = var_212_epsilon_0, x = var_211_cast_fp16)[name = string("op_212_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_213_cast_fp16 = mul(x = y_9_cast_fp16, y = var_212_cast_fp16)[name = string("op_213_cast_fp16")];
            tensor<fp16, [1024]> dd_trunk_layers_1_operator_norm_weight_to_fp16 = const()[name = string("dd_trunk_layers_1_operator_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72983168)))];
            tensor<fp16, [1, 12, 1024]> input_11_cast_fp16 = mul(x = dd_trunk_layers_1_operator_norm_weight_to_fp16, y = var_213_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1024, 1024]> dd_trunk_layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72985280)))];
            tensor<fp16, [1, 12, 1024]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = dd_trunk_layers_1_self_attn_q_proj_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> concat_8 = const()[name = string("concat_8"), val = tensor<int32, [4]>([1, 12, 32, 32])];
            tensor<fp16, [1, 12, 32, 32]> x_21_cast_fp16 = reshape(shape = concat_8, x = linear_8_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> var_227_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_227_cast_fp16")];
            tensor<int32, [1]> var_229_axes_0 = const()[name = string("op_229_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 32, 1]> var_229_cast_fp16 = reduce_max(axes = var_229_axes_0, keep_dims = var_26, x = var_227_cast_fp16)[name = string("op_229_cast_fp16")];
            tensor<fp16, [1, 12, 32, 1]> s_11_cast_fp16 = maximum(x = var_229_cast_fp16, y = var_25_to_fp16)[name = string("s_11_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> y_11_cast_fp16 = real_div(x = x_21_cast_fp16, y = s_11_cast_fp16)[name = string("y_11_cast_fp16")];
            fp16 var_24_promoted_5_to_fp16 = const()[name = string("op_24_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 32, 32]> var_232_cast_fp16 = pow(x = y_11_cast_fp16, y = var_24_promoted_5_to_fp16)[name = string("op_232_cast_fp16")];
            tensor<int32, [1]> var_234_axes_0 = const()[name = string("op_234_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_234_keep_dims_0 = const()[name = string("op_234_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 32, 1]> var_234_cast_fp16 = reduce_mean(axes = var_234_axes_0, keep_dims = var_234_keep_dims_0, x = var_232_cast_fp16)[name = string("op_234_cast_fp16")];
            fp32 var_235_epsilon_0 = const()[name = string("op_235_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 12, 32, 1]> var_235_cast_fp16 = inverse(epsilon = var_235_epsilon_0, x = s_11_cast_fp16)[name = string("op_235_cast_fp16")];
            fp16 var_236_to_fp16 = const()[name = string("op_236_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 32, 1]> var_237_cast_fp16 = mul(x = var_235_cast_fp16, y = var_236_to_fp16)[name = string("op_237_cast_fp16")];
            tensor<fp16, [1, 12, 32, 1]> var_238_cast_fp16 = real_div(x = var_237_cast_fp16, y = s_11_cast_fp16)[name = string("op_238_cast_fp16")];
            tensor<fp16, [1, 12, 32, 1]> var_239_cast_fp16 = add(x = var_234_cast_fp16, y = var_238_cast_fp16)[name = string("op_239_cast_fp16")];
            fp32 var_240_epsilon_0 = const()[name = string("op_240_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 32, 1]> var_240_cast_fp16 = rsqrt(epsilon = var_240_epsilon_0, x = var_239_cast_fp16)[name = string("op_240_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> var_241_cast_fp16 = mul(x = y_11_cast_fp16, y = var_240_cast_fp16)[name = string("op_241_cast_fp16")];
            tensor<fp16, [32]> dd_trunk_layers_1_self_attn_q_norm_weight_to_fp16 = const()[name = string("dd_trunk_layers_1_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75082496)))];
            tensor<fp16, [1, 12, 32, 32]> var_242_cast_fp16 = mul(x = dd_trunk_layers_1_self_attn_q_norm_weight_to_fp16, y = var_241_cast_fp16)[name = string("op_242_cast_fp16")];
            tensor<fp16, [256, 1024]> dd_trunk_layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75082624)))];
            tensor<fp16, [1, 12, 256]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = dd_trunk_layers_1_self_attn_k_proj_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> concat_9 = const()[name = string("concat_9"), val = tensor<int32, [4]>([1, 12, 8, 32])];
            tensor<fp16, [1, 12, 8, 32]> x_23_cast_fp16 = reshape(shape = concat_9, x = linear_9_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [1, 12, 8, 32]> var_249_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_249_cast_fp16")];
            tensor<int32, [1]> var_251_axes_0 = const()[name = string("op_251_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 8, 1]> var_251_cast_fp16 = reduce_max(axes = var_251_axes_0, keep_dims = var_26, x = var_249_cast_fp16)[name = string("op_251_cast_fp16")];
            tensor<fp16, [1, 12, 8, 1]> s_13_cast_fp16 = maximum(x = var_251_cast_fp16, y = var_25_to_fp16)[name = string("s_13_cast_fp16")];
            tensor<fp16, [1, 12, 8, 32]> y_13_cast_fp16 = real_div(x = x_23_cast_fp16, y = s_13_cast_fp16)[name = string("y_13_cast_fp16")];
            fp16 var_24_promoted_6_to_fp16 = const()[name = string("op_24_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 8, 32]> var_254_cast_fp16 = pow(x = y_13_cast_fp16, y = var_24_promoted_6_to_fp16)[name = string("op_254_cast_fp16")];
            tensor<int32, [1]> var_256_axes_0 = const()[name = string("op_256_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_256_keep_dims_0 = const()[name = string("op_256_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 8, 1]> var_256_cast_fp16 = reduce_mean(axes = var_256_axes_0, keep_dims = var_256_keep_dims_0, x = var_254_cast_fp16)[name = string("op_256_cast_fp16")];
            fp32 var_257_epsilon_0 = const()[name = string("op_257_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 12, 8, 1]> var_257_cast_fp16 = inverse(epsilon = var_257_epsilon_0, x = s_13_cast_fp16)[name = string("op_257_cast_fp16")];
            fp16 var_258_to_fp16 = const()[name = string("op_258_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 8, 1]> var_259_cast_fp16 = mul(x = var_257_cast_fp16, y = var_258_to_fp16)[name = string("op_259_cast_fp16")];
            tensor<fp16, [1, 12, 8, 1]> var_260_cast_fp16 = real_div(x = var_259_cast_fp16, y = s_13_cast_fp16)[name = string("op_260_cast_fp16")];
            tensor<fp16, [1, 12, 8, 1]> var_261_cast_fp16 = add(x = var_256_cast_fp16, y = var_260_cast_fp16)[name = string("op_261_cast_fp16")];
            fp32 var_262_epsilon_0 = const()[name = string("op_262_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 8, 1]> var_262_cast_fp16 = rsqrt(epsilon = var_262_epsilon_0, x = var_261_cast_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [1, 12, 8, 32]> var_263_cast_fp16 = mul(x = y_13_cast_fp16, y = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            tensor<fp16, [32]> dd_trunk_layers_1_self_attn_k_norm_weight_to_fp16 = const()[name = string("dd_trunk_layers_1_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75606976)))];
            tensor<fp16, [1, 12, 8, 32]> var_264_cast_fp16 = mul(x = dd_trunk_layers_1_self_attn_k_norm_weight_to_fp16, y = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<int32, [4]> x_27_perm_0 = const()[name = string("x_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 1024]> dd_trunk_layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75607104)))];
            tensor<fp16, [1, 12, 256]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = dd_trunk_layers_1_self_attn_v_proj_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [4]> concat_10 = const()[name = string("concat_10"), val = tensor<int32, [4]>([1, 12, 8, 32])];
            tensor<fp16, [1, 12, 8, 32]> var_269_cast_fp16 = reshape(shape = concat_10, x = linear_10_cast_fp16)[name = string("op_269_cast_fp16")];
            tensor<int32, [4]> v_5_perm_0 = const()[name = string("v_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 32, 32]> var_271_cast_fp16 = mul(x = var_242_cast_fp16, y = const_8_to_fp16)[name = string("op_271_cast_fp16")];
            tensor<int32, [2]> var_272_split_sizes_0 = const()[name = string("op_272_split_sizes_0"), val = tensor<int32, [2]>([16, 16])];
            int32 const_13 = const()[name = string("const_13"), val = int32(3)];
            tensor<fp16, [1, 12, 32, 16]> var_272_cast_fp16_0, tensor<fp16, [1, 12, 32, 16]> var_272_cast_fp16_1 = split(axis = const_13, split_sizes = var_272_split_sizes_0, x = var_242_cast_fp16)[name = string("op_272_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, 32, 16]> var_274_cast_fp16 = mul(x = var_272_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_274_cast_fp16")];
            bool var_276_interleave_0 = const()[name = string("op_276_interleave_0"), val = bool(false)];
            int32 const_14 = const()[name = string("const_14"), val = int32(3)];
            tensor<fp16, [1, 12, 32, 32]> var_276_cast_fp16 = concat(axis = const_14, interleave = var_276_interleave_0, values = (var_274_cast_fp16, var_272_cast_fp16_0))[name = string("op_276_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> var_277_cast_fp16 = mul(x = var_276_cast_fp16, y = const_11_to_fp16)[name = string("op_277_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> q_3_cast_fp16 = add(x = var_271_cast_fp16, y = var_277_cast_fp16)[name = string("q_3_cast_fp16")];
            tensor<fp16, [1, 8, 12, 32]> x_27_cast_fp16 = transpose(perm = x_27_perm_0, x = var_264_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 8, 12, 32]> var_279_cast_fp16 = mul(x = x_27_cast_fp16, y = const_0_to_fp16)[name = string("op_279_cast_fp16")];
            tensor<int32, [2]> var_280_split_sizes_0 = const()[name = string("op_280_split_sizes_0"), val = tensor<int32, [2]>([16, 16])];
            int32 var_280_axis_0 = const()[name = string("op_280_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 12, 16]> var_280_cast_fp16_0, tensor<fp16, [1, 8, 12, 16]> var_280_cast_fp16_1 = split(axis = var_280_axis_0, split_sizes = var_280_split_sizes_0, x = x_27_cast_fp16)[name = string("op_280_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 12, 16]> var_282_cast_fp16 = mul(x = var_280_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_282_cast_fp16")];
            bool var_284_interleave_0 = const()[name = string("op_284_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 12, 32]> var_284_cast_fp16 = concat(axis = var_27, interleave = var_284_interleave_0, values = (var_282_cast_fp16, var_280_cast_fp16_0))[name = string("op_284_cast_fp16")];
            tensor<fp16, [1, 8, 12, 32]> var_285_cast_fp16 = mul(x = var_284_cast_fp16, y = const_1_to_fp16)[name = string("op_285_cast_fp16")];
            tensor<fp16, [1, 8, 12, 32]> k_5_cast_fp16 = add(x = var_279_cast_fp16, y = var_285_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<int32, [1]> var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 12, 32]> var_287_cast_fp16 = expand_dims(axes = var_287_axes_0, x = k_5_cast_fp16)[name = string("op_287_cast_fp16")];
            tensor<int32, [5]> var_289_reps_0 = const()[name = string("op_289_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 8, 4, 12, 32]> var_289_cast_fp16 = tile(reps = var_289_reps_0, x = var_287_cast_fp16)[name = string("op_289_cast_fp16")];
            tensor<int32, [4]> concat_12 = const()[name = string("concat_12"), val = tensor<int32, [4]>([1, 32, 12, 32])];
            tensor<fp16, [1, 32, 12, 32]> k_7_cast_fp16 = reshape(shape = concat_12, x = var_289_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<int32, [1]> var_291_axes_0 = const()[name = string("op_291_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 12, 32]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = var_269_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 8, 1, 12, 32]> var_291_cast_fp16 = expand_dims(axes = var_291_axes_0, x = v_5_cast_fp16)[name = string("op_291_cast_fp16")];
            tensor<int32, [5]> var_293_reps_0 = const()[name = string("op_293_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 8, 4, 12, 32]> var_293_cast_fp16 = tile(reps = var_293_reps_0, x = var_291_cast_fp16)[name = string("op_293_cast_fp16")];
            tensor<int32, [4]> concat_14 = const()[name = string("concat_14"), val = tensor<int32, [4]>([1, 32, 12, 32])];
            tensor<fp16, [1, 32, 12, 32]> v_7_cast_fp16 = reshape(shape = concat_14, x = var_293_cast_fp16)[name = string("v_7_cast_fp16")];
            tensor<int32, [4]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_296_transpose_x_1 = const()[name = string("op_296_transpose_x_1"), val = bool(false)];
            bool var_296_transpose_y_1 = const()[name = string("op_296_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 32, 12, 32]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = q_3_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 32, 12, 12]> var_296_cast_fp16 = matmul(transpose_x = var_296_transpose_x_1, transpose_y = var_296_transpose_y_1, x = transpose_2_cast_fp16, y = k_7_cast_fp16)[name = string("op_296_cast_fp16")];
            fp16 var_297_to_fp16 = const()[name = string("op_297_to_fp16"), val = fp16(0x1.6ap-3)];
            tensor<fp16, [1, 32, 12, 12]> x_29_cast_fp16 = mul(x = var_296_cast_fp16, y = var_297_to_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [1]> var_300_axes_0 = const()[name = string("op_300_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 12, 1]> var_300_cast_fp16 = reduce_max(axes = var_300_axes_0, keep_dims = var_26, x = x_29_cast_fp16)[name = string("op_300_cast_fp16")];
            tensor<fp16, [1, 32, 12, 12]> x_31_cast_fp16 = sub(x = x_29_cast_fp16, y = var_300_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1, 32, 12, 12]> x_33_cast_fp16 = exp(x = x_31_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<int32, [1]> var_304_axes_0 = const()[name = string("op_304_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_304_keep_dims_0 = const()[name = string("op_304_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 12, 1]> var_304_cast_fp16 = reduce_sum(axes = var_304_axes_0, keep_dims = var_304_keep_dims_0, x = x_33_cast_fp16)[name = string("op_304_cast_fp16")];
            tensor<fp16, [1, 32, 12, 12]> var_305_cast_fp16 = real_div(x = x_33_cast_fp16, y = var_304_cast_fp16)[name = string("op_305_cast_fp16")];
            bool o_7_transpose_x_0 = const()[name = string("o_7_transpose_x_0"), val = bool(false)];
            bool o_7_transpose_y_0 = const()[name = string("o_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 32, 12, 32]> o_7_cast_fp16 = matmul(transpose_x = o_7_transpose_x_0, transpose_y = o_7_transpose_y_0, x = var_305_cast_fp16, y = v_7_cast_fp16)[name = string("o_7_cast_fp16")];
            tensor<int32, [4]> o_9_perm_0 = const()[name = string("o_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [32, 1024]> dd_trunk_layers_1_self_attn_gate_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_1_self_attn_gate_proj_weight_to_fp16"), val = tensor<fp16, [32, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76131456)))];
            tensor<fp16, [32]> linear_11_bias_0_to_fp16 = const()[name = string("linear_11_bias_0_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53576320)))];
            tensor<fp16, [1, 12, 32]> linear_11_cast_fp16 = linear(bias = linear_11_bias_0_to_fp16, weight = dd_trunk_layers_1_self_attn_gate_proj_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 12, 32]> var_311_cast_fp16 = sigmoid(x = linear_11_cast_fp16)[name = string("op_311_cast_fp16")];
            tensor<int32, [1]> var_312_axes_0 = const()[name = string("op_312_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 32, 1]> var_312_cast_fp16 = expand_dims(axes = var_312_axes_0, x = var_311_cast_fp16)[name = string("op_312_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> o_9_cast_fp16 = transpose(perm = o_9_perm_0, x = o_7_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 12, 32, 32]> o_11_cast_fp16 = mul(x = o_9_cast_fp16, y = var_312_cast_fp16)[name = string("o_11_cast_fp16")];
            tensor<int32, [3]> concat_15 = const()[name = string("concat_15"), val = tensor<int32, [3]>([1, 12, 1024])];
            tensor<fp16, [1, 12, 1024]> input_13_cast_fp16 = reshape(shape = concat_15, x = o_11_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 1024]> dd_trunk_layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76197056)))];
            tensor<fp16, [1, 12, 1024]> linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = dd_trunk_layers_1_self_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> x_37_cast_fp16 = add(x = x_19_cast_fp16, y = linear_12_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_319_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_319_cast_fp16")];
            tensor<int32, [1]> var_321_axes_0 = const()[name = string("op_321_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 1]> var_321_cast_fp16 = reduce_max(axes = var_321_axes_0, keep_dims = var_26, x = var_319_cast_fp16)[name = string("op_321_cast_fp16")];
            tensor<fp16, [1, 12, 1]> s_15_cast_fp16 = maximum(x = var_321_cast_fp16, y = var_25_to_fp16)[name = string("s_15_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> y_15_cast_fp16 = real_div(x = x_37_cast_fp16, y = s_15_cast_fp16)[name = string("y_15_cast_fp16")];
            fp16 var_24_promoted_7_to_fp16 = const()[name = string("op_24_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 1024]> var_324_cast_fp16 = pow(x = y_15_cast_fp16, y = var_24_promoted_7_to_fp16)[name = string("op_324_cast_fp16")];
            tensor<int32, [1]> var_326_axes_0 = const()[name = string("op_326_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_326_keep_dims_0 = const()[name = string("op_326_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 1]> var_326_cast_fp16 = reduce_mean(axes = var_326_axes_0, keep_dims = var_326_keep_dims_0, x = var_324_cast_fp16)[name = string("op_326_cast_fp16")];
            fp32 var_327_epsilon_0 = const()[name = string("op_327_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 12, 1]> var_327_cast_fp16 = inverse(epsilon = var_327_epsilon_0, x = s_15_cast_fp16)[name = string("op_327_cast_fp16")];
            fp16 var_328_to_fp16 = const()[name = string("op_328_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 1]> var_329_cast_fp16 = mul(x = var_327_cast_fp16, y = var_328_to_fp16)[name = string("op_329_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_330_cast_fp16 = real_div(x = var_329_cast_fp16, y = s_15_cast_fp16)[name = string("op_330_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_331_cast_fp16 = add(x = var_326_cast_fp16, y = var_330_cast_fp16)[name = string("op_331_cast_fp16")];
            fp32 var_332_epsilon_0 = const()[name = string("op_332_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 1]> var_332_cast_fp16 = rsqrt(epsilon = var_332_epsilon_0, x = var_331_cast_fp16)[name = string("op_332_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_333_cast_fp16 = mul(x = y_15_cast_fp16, y = var_332_cast_fp16)[name = string("op_333_cast_fp16")];
            tensor<fp16, [1024]> dd_trunk_layers_1_ffn_norm_weight_to_fp16 = const()[name = string("dd_trunk_layers_1_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78294272)))];
            tensor<fp16, [1, 12, 1024]> input_15_cast_fp16 = mul(x = dd_trunk_layers_1_ffn_norm_weight_to_fp16, y = var_333_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [2816, 1024]> dd_trunk_layers_1_feed_forward_w1_weight_to_fp16 = const()[name = string("dd_trunk_layers_1_feed_forward_w1_weight_to_fp16"), val = tensor<fp16, [2816, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78296384)))];
            tensor<fp16, [1, 12, 2816]> linear_13_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = dd_trunk_layers_1_feed_forward_w1_weight_to_fp16, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [2816, 1024]> dd_trunk_layers_1_feed_forward_w3_weight_to_fp16 = const()[name = string("dd_trunk_layers_1_feed_forward_w3_weight_to_fp16"), val = tensor<fp16, [2816, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84063616)))];
            tensor<fp16, [1, 12, 2816]> linear_14_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = dd_trunk_layers_1_feed_forward_w3_weight_to_fp16, x = input_15_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<fp16, [1, 12, 2816]> var_342_cast_fp16 = silu(x = linear_13_cast_fp16)[name = string("op_342_cast_fp16")];
            tensor<fp16, [1, 12, 2816]> input_19_cast_fp16 = mul(x = var_342_cast_fp16, y = linear_14_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1024, 2816]> dd_trunk_layers_1_feed_forward_w2_weight_to_fp16 = const()[name = string("dd_trunk_layers_1_feed_forward_w2_weight_to_fp16"), val = tensor<fp16, [1024, 2816]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89830848)))];
            tensor<fp16, [1, 12, 1024]> linear_15_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = dd_trunk_layers_1_feed_forward_w2_weight_to_fp16, x = input_19_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> x_39_cast_fp16 = add(x = x_37_cast_fp16, y = linear_15_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_352_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_352_cast_fp16")];
            tensor<int32, [1]> var_354_axes_0 = const()[name = string("op_354_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 1]> var_354_cast_fp16 = reduce_max(axes = var_354_axes_0, keep_dims = var_26, x = var_352_cast_fp16)[name = string("op_354_cast_fp16")];
            tensor<fp16, [1, 12, 1]> s_17_cast_fp16 = maximum(x = var_354_cast_fp16, y = var_25_to_fp16)[name = string("s_17_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> y_17_cast_fp16 = real_div(x = x_39_cast_fp16, y = s_17_cast_fp16)[name = string("y_17_cast_fp16")];
            fp16 var_24_promoted_8_to_fp16 = const()[name = string("op_24_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 1024]> var_357_cast_fp16 = pow(x = y_17_cast_fp16, y = var_24_promoted_8_to_fp16)[name = string("op_357_cast_fp16")];
            tensor<int32, [1]> var_359_axes_0 = const()[name = string("op_359_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_359_keep_dims_0 = const()[name = string("op_359_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 1]> var_359_cast_fp16 = reduce_mean(axes = var_359_axes_0, keep_dims = var_359_keep_dims_0, x = var_357_cast_fp16)[name = string("op_359_cast_fp16")];
            fp32 var_360_epsilon_0 = const()[name = string("op_360_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 12, 1]> var_360_cast_fp16 = inverse(epsilon = var_360_epsilon_0, x = s_17_cast_fp16)[name = string("op_360_cast_fp16")];
            fp16 var_361_to_fp16 = const()[name = string("op_361_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 1]> var_362_cast_fp16 = mul(x = var_360_cast_fp16, y = var_361_to_fp16)[name = string("op_362_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_363_cast_fp16 = real_div(x = var_362_cast_fp16, y = s_17_cast_fp16)[name = string("op_363_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_364_cast_fp16 = add(x = var_359_cast_fp16, y = var_363_cast_fp16)[name = string("op_364_cast_fp16")];
            fp32 var_365_epsilon_0 = const()[name = string("op_365_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 1]> var_365_cast_fp16 = rsqrt(epsilon = var_365_epsilon_0, x = var_364_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_366_cast_fp16 = mul(x = y_17_cast_fp16, y = var_365_cast_fp16)[name = string("op_366_cast_fp16")];
            tensor<fp16, [1024]> dd_trunk_layers_2_operator_norm_weight_to_fp16 = const()[name = string("dd_trunk_layers_2_operator_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95598080)))];
            tensor<fp16, [1, 12, 1024]> input_21_cast_fp16 = mul(x = dd_trunk_layers_2_operator_norm_weight_to_fp16, y = var_366_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1024, 1024]> dd_trunk_layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95600192)))];
            tensor<fp16, [1, 12, 1024]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = dd_trunk_layers_2_self_attn_q_proj_weight_to_fp16, x = input_21_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> concat_16 = const()[name = string("concat_16"), val = tensor<int32, [4]>([1, 12, 32, 32])];
            tensor<fp16, [1, 12, 32, 32]> x_41_cast_fp16 = reshape(shape = concat_16, x = linear_16_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> var_380_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_380_cast_fp16")];
            tensor<int32, [1]> var_382_axes_0 = const()[name = string("op_382_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 32, 1]> var_382_cast_fp16 = reduce_max(axes = var_382_axes_0, keep_dims = var_26, x = var_380_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<fp16, [1, 12, 32, 1]> s_19_cast_fp16 = maximum(x = var_382_cast_fp16, y = var_25_to_fp16)[name = string("s_19_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> y_19_cast_fp16 = real_div(x = x_41_cast_fp16, y = s_19_cast_fp16)[name = string("y_19_cast_fp16")];
            fp16 var_24_promoted_9_to_fp16 = const()[name = string("op_24_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 32, 32]> var_385_cast_fp16 = pow(x = y_19_cast_fp16, y = var_24_promoted_9_to_fp16)[name = string("op_385_cast_fp16")];
            tensor<int32, [1]> var_387_axes_0 = const()[name = string("op_387_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_387_keep_dims_0 = const()[name = string("op_387_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 32, 1]> var_387_cast_fp16 = reduce_mean(axes = var_387_axes_0, keep_dims = var_387_keep_dims_0, x = var_385_cast_fp16)[name = string("op_387_cast_fp16")];
            fp32 var_388_epsilon_0 = const()[name = string("op_388_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 12, 32, 1]> var_388_cast_fp16 = inverse(epsilon = var_388_epsilon_0, x = s_19_cast_fp16)[name = string("op_388_cast_fp16")];
            fp16 var_389_to_fp16 = const()[name = string("op_389_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 32, 1]> var_390_cast_fp16 = mul(x = var_388_cast_fp16, y = var_389_to_fp16)[name = string("op_390_cast_fp16")];
            tensor<fp16, [1, 12, 32, 1]> var_391_cast_fp16 = real_div(x = var_390_cast_fp16, y = s_19_cast_fp16)[name = string("op_391_cast_fp16")];
            tensor<fp16, [1, 12, 32, 1]> var_392_cast_fp16 = add(x = var_387_cast_fp16, y = var_391_cast_fp16)[name = string("op_392_cast_fp16")];
            fp32 var_393_epsilon_0 = const()[name = string("op_393_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 32, 1]> var_393_cast_fp16 = rsqrt(epsilon = var_393_epsilon_0, x = var_392_cast_fp16)[name = string("op_393_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> var_394_cast_fp16 = mul(x = y_19_cast_fp16, y = var_393_cast_fp16)[name = string("op_394_cast_fp16")];
            tensor<fp16, [32]> dd_trunk_layers_2_self_attn_q_norm_weight_to_fp16 = const()[name = string("dd_trunk_layers_2_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97697408)))];
            tensor<fp16, [1, 12, 32, 32]> var_395_cast_fp16 = mul(x = dd_trunk_layers_2_self_attn_q_norm_weight_to_fp16, y = var_394_cast_fp16)[name = string("op_395_cast_fp16")];
            tensor<fp16, [256, 1024]> dd_trunk_layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97697536)))];
            tensor<fp16, [1, 12, 256]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = dd_trunk_layers_2_self_attn_k_proj_weight_to_fp16, x = input_21_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<int32, [4]> concat_17 = const()[name = string("concat_17"), val = tensor<int32, [4]>([1, 12, 8, 32])];
            tensor<fp16, [1, 12, 8, 32]> x_43_cast_fp16 = reshape(shape = concat_17, x = linear_17_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1, 12, 8, 32]> var_402_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_402_cast_fp16")];
            tensor<int32, [1]> var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 8, 1]> var_404_cast_fp16 = reduce_max(axes = var_404_axes_0, keep_dims = var_26, x = var_402_cast_fp16)[name = string("op_404_cast_fp16")];
            tensor<fp16, [1, 12, 8, 1]> s_21_cast_fp16 = maximum(x = var_404_cast_fp16, y = var_25_to_fp16)[name = string("s_21_cast_fp16")];
            tensor<fp16, [1, 12, 8, 32]> y_21_cast_fp16 = real_div(x = x_43_cast_fp16, y = s_21_cast_fp16)[name = string("y_21_cast_fp16")];
            fp16 var_24_promoted_10_to_fp16 = const()[name = string("op_24_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 8, 32]> var_407_cast_fp16 = pow(x = y_21_cast_fp16, y = var_24_promoted_10_to_fp16)[name = string("op_407_cast_fp16")];
            tensor<int32, [1]> var_409_axes_0 = const()[name = string("op_409_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_409_keep_dims_0 = const()[name = string("op_409_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 8, 1]> var_409_cast_fp16 = reduce_mean(axes = var_409_axes_0, keep_dims = var_409_keep_dims_0, x = var_407_cast_fp16)[name = string("op_409_cast_fp16")];
            fp32 var_410_epsilon_0 = const()[name = string("op_410_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 12, 8, 1]> var_410_cast_fp16 = inverse(epsilon = var_410_epsilon_0, x = s_21_cast_fp16)[name = string("op_410_cast_fp16")];
            fp16 var_411_to_fp16 = const()[name = string("op_411_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 8, 1]> var_412_cast_fp16 = mul(x = var_410_cast_fp16, y = var_411_to_fp16)[name = string("op_412_cast_fp16")];
            tensor<fp16, [1, 12, 8, 1]> var_413_cast_fp16 = real_div(x = var_412_cast_fp16, y = s_21_cast_fp16)[name = string("op_413_cast_fp16")];
            tensor<fp16, [1, 12, 8, 1]> var_414_cast_fp16 = add(x = var_409_cast_fp16, y = var_413_cast_fp16)[name = string("op_414_cast_fp16")];
            fp32 var_415_epsilon_0 = const()[name = string("op_415_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 8, 1]> var_415_cast_fp16 = rsqrt(epsilon = var_415_epsilon_0, x = var_414_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<fp16, [1, 12, 8, 32]> var_416_cast_fp16 = mul(x = y_21_cast_fp16, y = var_415_cast_fp16)[name = string("op_416_cast_fp16")];
            tensor<fp16, [32]> dd_trunk_layers_2_self_attn_k_norm_weight_to_fp16 = const()[name = string("dd_trunk_layers_2_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98221888)))];
            tensor<fp16, [1, 12, 8, 32]> var_417_cast_fp16 = mul(x = dd_trunk_layers_2_self_attn_k_norm_weight_to_fp16, y = var_416_cast_fp16)[name = string("op_417_cast_fp16")];
            tensor<int32, [4]> x_47_perm_0 = const()[name = string("x_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 1024]> dd_trunk_layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98222016)))];
            tensor<fp16, [1, 12, 256]> linear_18_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = dd_trunk_layers_2_self_attn_v_proj_weight_to_fp16, x = input_21_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [4]> concat_18 = const()[name = string("concat_18"), val = tensor<int32, [4]>([1, 12, 8, 32])];
            tensor<fp16, [1, 12, 8, 32]> var_422_cast_fp16 = reshape(shape = concat_18, x = linear_18_cast_fp16)[name = string("op_422_cast_fp16")];
            tensor<int32, [4]> v_9_perm_0 = const()[name = string("v_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 32, 32]> var_424_cast_fp16 = mul(x = var_395_cast_fp16, y = const_8_to_fp16)[name = string("op_424_cast_fp16")];
            tensor<int32, [2]> var_425_split_sizes_0 = const()[name = string("op_425_split_sizes_0"), val = tensor<int32, [2]>([16, 16])];
            int32 const_17 = const()[name = string("const_17"), val = int32(3)];
            tensor<fp16, [1, 12, 32, 16]> var_425_cast_fp16_0, tensor<fp16, [1, 12, 32, 16]> var_425_cast_fp16_1 = split(axis = const_17, split_sizes = var_425_split_sizes_0, x = var_395_cast_fp16)[name = string("op_425_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, 32, 16]> var_427_cast_fp16 = mul(x = var_425_cast_fp16_1, y = const_6_promoted_to_fp16)[name = string("op_427_cast_fp16")];
            bool var_429_interleave_0 = const()[name = string("op_429_interleave_0"), val = bool(false)];
            int32 const_18 = const()[name = string("const_18"), val = int32(3)];
            tensor<fp16, [1, 12, 32, 32]> var_429_cast_fp16 = concat(axis = const_18, interleave = var_429_interleave_0, values = (var_427_cast_fp16, var_425_cast_fp16_0))[name = string("op_429_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> var_430_cast_fp16 = mul(x = var_429_cast_fp16, y = const_11_to_fp16)[name = string("op_430_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> q_cast_fp16 = add(x = var_424_cast_fp16, y = var_430_cast_fp16)[name = string("q_cast_fp16")];
            tensor<fp16, [1, 8, 12, 32]> x_47_cast_fp16 = transpose(perm = x_47_perm_0, x = var_417_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 8, 12, 32]> var_432_cast_fp16 = mul(x = x_47_cast_fp16, y = const_0_to_fp16)[name = string("op_432_cast_fp16")];
            tensor<int32, [2]> var_433_split_sizes_0 = const()[name = string("op_433_split_sizes_0"), val = tensor<int32, [2]>([16, 16])];
            int32 var_433_axis_0 = const()[name = string("op_433_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 12, 16]> var_433_cast_fp16_0, tensor<fp16, [1, 8, 12, 16]> var_433_cast_fp16_1 = split(axis = var_433_axis_0, split_sizes = var_433_split_sizes_0, x = x_47_cast_fp16)[name = string("op_433_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 12, 16]> var_435_cast_fp16 = mul(x = var_433_cast_fp16_1, y = const_7_promoted_to_fp16)[name = string("op_435_cast_fp16")];
            bool var_437_interleave_0 = const()[name = string("op_437_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 12, 32]> var_437_cast_fp16 = concat(axis = var_27, interleave = var_437_interleave_0, values = (var_435_cast_fp16, var_433_cast_fp16_0))[name = string("op_437_cast_fp16")];
            tensor<fp16, [1, 8, 12, 32]> var_438_cast_fp16 = mul(x = var_437_cast_fp16, y = const_1_to_fp16)[name = string("op_438_cast_fp16")];
            tensor<fp16, [1, 8, 12, 32]> k_9_cast_fp16 = add(x = var_432_cast_fp16, y = var_438_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [1]> var_440_axes_0 = const()[name = string("op_440_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 12, 32]> var_440_cast_fp16 = expand_dims(axes = var_440_axes_0, x = k_9_cast_fp16)[name = string("op_440_cast_fp16")];
            tensor<int32, [5]> var_442_reps_0 = const()[name = string("op_442_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 8, 4, 12, 32]> var_442_cast_fp16 = tile(reps = var_442_reps_0, x = var_440_cast_fp16)[name = string("op_442_cast_fp16")];
            tensor<int32, [4]> concat_20 = const()[name = string("concat_20"), val = tensor<int32, [4]>([1, 32, 12, 32])];
            tensor<fp16, [1, 32, 12, 32]> k_cast_fp16 = reshape(shape = concat_20, x = var_442_cast_fp16)[name = string("k_cast_fp16")];
            tensor<int32, [1]> var_444_axes_0 = const()[name = string("op_444_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 12, 32]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_422_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 8, 1, 12, 32]> var_444_cast_fp16 = expand_dims(axes = var_444_axes_0, x = v_9_cast_fp16)[name = string("op_444_cast_fp16")];
            tensor<int32, [5]> var_446_reps_0 = const()[name = string("op_446_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 8, 4, 12, 32]> var_446_cast_fp16 = tile(reps = var_446_reps_0, x = var_444_cast_fp16)[name = string("op_446_cast_fp16")];
            tensor<int32, [4]> concat_22 = const()[name = string("concat_22"), val = tensor<int32, [4]>([1, 32, 12, 32])];
            tensor<fp16, [1, 32, 12, 32]> v_cast_fp16 = reshape(shape = concat_22, x = var_446_cast_fp16)[name = string("v_cast_fp16")];
            tensor<int32, [4]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_449_transpose_x_1 = const()[name = string("op_449_transpose_x_1"), val = bool(false)];
            bool var_449_transpose_y_1 = const()[name = string("op_449_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 32, 12, 32]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = q_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 32, 12, 12]> var_449_cast_fp16 = matmul(transpose_x = var_449_transpose_x_1, transpose_y = var_449_transpose_y_1, x = transpose_4_cast_fp16, y = k_cast_fp16)[name = string("op_449_cast_fp16")];
            fp16 var_450_to_fp16 = const()[name = string("op_450_to_fp16"), val = fp16(0x1.6ap-3)];
            tensor<fp16, [1, 32, 12, 12]> x_49_cast_fp16 = mul(x = var_449_cast_fp16, y = var_450_to_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [1]> var_453_axes_0 = const()[name = string("op_453_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 12, 1]> var_453_cast_fp16 = reduce_max(axes = var_453_axes_0, keep_dims = var_26, x = x_49_cast_fp16)[name = string("op_453_cast_fp16")];
            tensor<fp16, [1, 32, 12, 12]> x_51_cast_fp16 = sub(x = x_49_cast_fp16, y = var_453_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [1, 32, 12, 12]> x_53_cast_fp16 = exp(x = x_51_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<int32, [1]> var_457_axes_0 = const()[name = string("op_457_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_457_keep_dims_0 = const()[name = string("op_457_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 12, 1]> var_457_cast_fp16 = reduce_sum(axes = var_457_axes_0, keep_dims = var_457_keep_dims_0, x = x_53_cast_fp16)[name = string("op_457_cast_fp16")];
            tensor<fp16, [1, 32, 12, 12]> var_458_cast_fp16 = real_div(x = x_53_cast_fp16, y = var_457_cast_fp16)[name = string("op_458_cast_fp16")];
            bool o_13_transpose_x_0 = const()[name = string("o_13_transpose_x_0"), val = bool(false)];
            bool o_13_transpose_y_0 = const()[name = string("o_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 32, 12, 32]> o_13_cast_fp16 = matmul(transpose_x = o_13_transpose_x_0, transpose_y = o_13_transpose_y_0, x = var_458_cast_fp16, y = v_cast_fp16)[name = string("o_13_cast_fp16")];
            tensor<int32, [4]> o_15_perm_0 = const()[name = string("o_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [32, 1024]> dd_trunk_layers_2_self_attn_gate_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_2_self_attn_gate_proj_weight_to_fp16"), val = tensor<fp16, [32, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98746368)))];
            tensor<fp16, [32]> linear_19_bias_0_to_fp16 = const()[name = string("linear_19_bias_0_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53576320)))];
            tensor<fp16, [1, 12, 32]> linear_19_cast_fp16 = linear(bias = linear_19_bias_0_to_fp16, weight = dd_trunk_layers_2_self_attn_gate_proj_weight_to_fp16, x = input_21_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 12, 32]> var_464_cast_fp16 = sigmoid(x = linear_19_cast_fp16)[name = string("op_464_cast_fp16")];
            tensor<int32, [1]> var_465_axes_0 = const()[name = string("op_465_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 32, 1]> var_465_cast_fp16 = expand_dims(axes = var_465_axes_0, x = var_464_cast_fp16)[name = string("op_465_cast_fp16")];
            tensor<fp16, [1, 12, 32, 32]> o_15_cast_fp16 = transpose(perm = o_15_perm_0, x = o_13_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 12, 32, 32]> o_cast_fp16 = mul(x = o_15_cast_fp16, y = var_465_cast_fp16)[name = string("o_cast_fp16")];
            tensor<int32, [3]> concat_23 = const()[name = string("concat_23"), val = tensor<int32, [3]>([1, 12, 1024])];
            tensor<fp16, [1, 12, 1024]> input_23_cast_fp16 = reshape(shape = concat_23, x = o_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1024, 1024]> dd_trunk_layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("dd_trunk_layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98811968)))];
            tensor<fp16, [1, 12, 1024]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = dd_trunk_layers_2_self_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> x_57_cast_fp16 = add(x = x_39_cast_fp16, y = linear_20_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_472_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_472_cast_fp16")];
            tensor<int32, [1]> var_474_axes_0 = const()[name = string("op_474_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 1]> var_474_cast_fp16 = reduce_max(axes = var_474_axes_0, keep_dims = var_26, x = var_472_cast_fp16)[name = string("op_474_cast_fp16")];
            tensor<fp16, [1, 12, 1]> s_23_cast_fp16 = maximum(x = var_474_cast_fp16, y = var_25_to_fp16)[name = string("s_23_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> y_23_cast_fp16 = real_div(x = x_57_cast_fp16, y = s_23_cast_fp16)[name = string("y_23_cast_fp16")];
            fp16 var_24_promoted_11_to_fp16 = const()[name = string("op_24_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 1024]> var_477_cast_fp16 = pow(x = y_23_cast_fp16, y = var_24_promoted_11_to_fp16)[name = string("op_477_cast_fp16")];
            tensor<int32, [1]> var_479_axes_0 = const()[name = string("op_479_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_479_keep_dims_0 = const()[name = string("op_479_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 1]> var_479_cast_fp16 = reduce_mean(axes = var_479_axes_0, keep_dims = var_479_keep_dims_0, x = var_477_cast_fp16)[name = string("op_479_cast_fp16")];
            fp32 var_480_epsilon_0 = const()[name = string("op_480_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 12, 1]> var_480_cast_fp16 = inverse(epsilon = var_480_epsilon_0, x = s_23_cast_fp16)[name = string("op_480_cast_fp16")];
            fp16 var_481_to_fp16 = const()[name = string("op_481_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 1]> var_482_cast_fp16 = mul(x = var_480_cast_fp16, y = var_481_to_fp16)[name = string("op_482_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_483_cast_fp16 = real_div(x = var_482_cast_fp16, y = s_23_cast_fp16)[name = string("op_483_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_484_cast_fp16 = add(x = var_479_cast_fp16, y = var_483_cast_fp16)[name = string("op_484_cast_fp16")];
            fp32 var_485_epsilon_0 = const()[name = string("op_485_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 1]> var_485_cast_fp16 = rsqrt(epsilon = var_485_epsilon_0, x = var_484_cast_fp16)[name = string("op_485_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_486_cast_fp16 = mul(x = y_23_cast_fp16, y = var_485_cast_fp16)[name = string("op_486_cast_fp16")];
            tensor<fp16, [1024]> dd_trunk_layers_2_ffn_norm_weight_to_fp16 = const()[name = string("dd_trunk_layers_2_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100909184)))];
            tensor<fp16, [1, 12, 1024]> input_25_cast_fp16 = mul(x = dd_trunk_layers_2_ffn_norm_weight_to_fp16, y = var_486_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [2816, 1024]> dd_trunk_layers_2_feed_forward_w1_weight_to_fp16 = const()[name = string("dd_trunk_layers_2_feed_forward_w1_weight_to_fp16"), val = tensor<fp16, [2816, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100911296)))];
            tensor<fp16, [1, 12, 2816]> linear_21_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = dd_trunk_layers_2_feed_forward_w1_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [2816, 1024]> dd_trunk_layers_2_feed_forward_w3_weight_to_fp16 = const()[name = string("dd_trunk_layers_2_feed_forward_w3_weight_to_fp16"), val = tensor<fp16, [2816, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106678528)))];
            tensor<fp16, [1, 12, 2816]> linear_22_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = dd_trunk_layers_2_feed_forward_w3_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [1, 12, 2816]> var_495_cast_fp16 = silu(x = linear_21_cast_fp16)[name = string("op_495_cast_fp16")];
            tensor<fp16, [1, 12, 2816]> input_cast_fp16 = mul(x = var_495_cast_fp16, y = linear_22_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [1024, 2816]> dd_trunk_layers_2_feed_forward_w2_weight_to_fp16 = const()[name = string("dd_trunk_layers_2_feed_forward_w2_weight_to_fp16"), val = tensor<fp16, [1024, 2816]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112445760)))];
            tensor<fp16, [1, 12, 1024]> linear_23_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = dd_trunk_layers_2_feed_forward_w2_weight_to_fp16, x = input_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> x_cast_fp16 = add(x = x_57_cast_fp16, y = linear_23_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_501_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_501_cast_fp16")];
            tensor<int32, [1]> var_503_axes_0 = const()[name = string("op_503_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 12, 1]> var_503_cast_fp16 = reduce_max(axes = var_503_axes_0, keep_dims = var_26, x = var_501_cast_fp16)[name = string("op_503_cast_fp16")];
            tensor<fp16, [1, 12, 1]> s_cast_fp16 = maximum(x = var_503_cast_fp16, y = var_25_to_fp16)[name = string("s_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> y_cast_fp16 = real_div(x = x_cast_fp16, y = s_cast_fp16)[name = string("y_cast_fp16")];
            fp16 var_24_promoted_12_to_fp16 = const()[name = string("op_24_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 1024]> var_506_cast_fp16 = pow(x = y_cast_fp16, y = var_24_promoted_12_to_fp16)[name = string("op_506_cast_fp16")];
            tensor<int32, [1]> var_508_axes_0 = const()[name = string("op_508_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_508_keep_dims_0 = const()[name = string("op_508_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 1]> var_508_cast_fp16 = reduce_mean(axes = var_508_axes_0, keep_dims = var_508_keep_dims_0, x = var_506_cast_fp16)[name = string("op_508_cast_fp16")];
            fp32 var_509_epsilon_0 = const()[name = string("op_509_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 12, 1]> var_509_cast_fp16 = inverse(epsilon = var_509_epsilon_0, x = s_cast_fp16)[name = string("op_509_cast_fp16")];
            fp16 var_510_to_fp16 = const()[name = string("op_510_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 1]> var_511_cast_fp16 = mul(x = var_509_cast_fp16, y = var_510_to_fp16)[name = string("op_511_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_512_cast_fp16 = real_div(x = var_511_cast_fp16, y = s_cast_fp16)[name = string("op_512_cast_fp16")];
            tensor<fp16, [1, 12, 1]> var_513_cast_fp16 = add(x = var_508_cast_fp16, y = var_512_cast_fp16)[name = string("op_513_cast_fp16")];
            fp32 var_514_epsilon_0 = const()[name = string("op_514_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 1]> var_514_cast_fp16 = rsqrt(epsilon = var_514_epsilon_0, x = var_513_cast_fp16)[name = string("op_514_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_515_cast_fp16 = mul(x = y_cast_fp16, y = var_514_cast_fp16)[name = string("op_515_cast_fp16")];
            tensor<fp16, [1024]> dd_trunk_embedding_norm_weight_to_fp16 = const()[name = string("dd_trunk_embedding_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118212992)))];
            tensor<fp16, [1, 12, 1024]> out_cast_fp16 = mul(x = dd_trunk_embedding_norm_weight_to_fp16, y = var_515_cast_fp16)[name = string("out_cast_fp16")];
            fp16 var_517_promoted_to_fp16 = const()[name = string("op_517_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 12, 1024]> var_518_cast_fp16 = pow(x = out_cast_fp16, y = var_517_promoted_to_fp16)[name = string("op_518_cast_fp16")];
            tensor<int32, [1]> var_523_axes_0 = const()[name = string("op_523_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_523_keep_dims_0 = const()[name = string("op_523_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 12, 1]> var_523_cast_fp16 = reduce_mean(axes = var_523_axes_0, keep_dims = var_523_keep_dims_0, x = var_518_cast_fp16)[name = string("op_523_cast_fp16")];
            fp16 var_525_to_fp16 = const()[name = string("op_525_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 12, 1]> var_526_cast_fp16 = add(x = var_523_cast_fp16, y = var_525_to_fp16)[name = string("op_526_cast_fp16")];
            fp32 var_527_epsilon_0 = const()[name = string("op_527_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 12, 1]> var_527_cast_fp16 = rsqrt(epsilon = var_527_epsilon_0, x = var_526_cast_fp16)[name = string("op_527_cast_fp16")];
            tensor<fp16, [1, 12, 1024]> var_528_cast_fp16 = mul(x = out_cast_fp16, y = var_527_cast_fp16)[name = string("op_528_cast_fp16")];
            tensor<fp16, [12, 1024]> dd_head_norm_to_fp16 = const()[name = string("dd_head_norm_to_fp16"), val = tensor<fp16, [12, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118215104)))];
            tensor<fp16, [1, 12, 1024]> xn_cast_fp16 = mul(x = var_528_cast_fp16, y = dd_head_norm_to_fp16)[name = string("xn_cast_fp16")];
            tensor<int32, [1]> var_531_axes_0 = const()[name = string("op_531_axes_0"), val = tensor<int32, [1]>([-2])];
            tensor<fp16, [1, 12, 1, 1024]> var_531_cast_fp16 = expand_dims(axes = var_531_axes_0, x = xn_cast_fp16)[name = string("op_531_cast_fp16")];
            bool var_535_transpose_x_0 = const()[name = string("op_535_transpose_x_0"), val = bool(false)];
            bool var_535_transpose_y_0 = const()[name = string("op_535_transpose_y_0"), val = bool(false)];
            tensor<fp16, [12, 1024, 2017]> var_534_to_fp16 = const()[name = string("op_534_to_fp16"), val = tensor<fp16, [12, 1024, 2017]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118239744)))];
            tensor<fp16, [1, 12, 1, 2017]> var_535_cast_fp16 = matmul(transpose_x = var_535_transpose_x_0, transpose_y = var_535_transpose_y_0, x = var_531_cast_fp16, y = var_534_to_fp16)[name = string("op_535_cast_fp16")];
            tensor<int32, [1]> var_537_axes_0 = const()[name = string("op_537_axes_0"), val = tensor<int32, [1]>([-2])];
            tensor<fp16, [1, 12, 2017]> logits = squeeze(axes = var_537_axes_0, x = var_535_cast_fp16)[name = string("op_537_cast_fp16")];
        } -> (logits);
}