program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-milinternal", ""}, {"coremltools-version", "9.0"}})]
{
    func length_1<ios18>(tensor<fp32, [1, 2048, 1, 1]> inputs_embeds, state<tensor<fp16, [9, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [1]> position_index_seed, state<tensor<fp16, [9, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725952))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17321280))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17308928))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17327488))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29922816))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29910464))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29929024))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42516160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42512000))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42518272))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243840))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243264))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47244160))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51442688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51438528))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444800))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64040128))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64027776))))[name = string("layers_1_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64046336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76633472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76629312))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76635584))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80834112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80829952))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80836224))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361152))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81360576))))[name = string("layers_2_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361472))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85560000))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85555840))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85562112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98145088))))[name = string("layers_2_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110758976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110746624))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110765184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123352320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123348160))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123354432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127552960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127548800))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127555072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128079424))))[name = string("layers_3_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080320))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140675648))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140663296))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140681856))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153277184))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153264832))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153283392))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165870528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165866368))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165872640))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170071168))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170067008))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170073280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170597632))))[name = string("layers_4_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174797056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174792896))))[name = string("layers_4_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174799168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187394496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187382144))))[name = string("layers_4_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187400704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199996032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199983680))))[name = string("layers_4_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_4_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200002240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212589376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212585216))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212591488))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216790016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216785856))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216792128))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317056))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217316480))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317376))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221515904))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221511744))))[name = string("layers_5_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221518016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234113344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234100992))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234119552))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246714880))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246702528))))[name = string("layers_5_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_5_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246721088))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259308224))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259304064))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259310336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263508864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263504704))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263510976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035328))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264036224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268234752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268230592))))[name = string("layers_6_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268236864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280832192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280819840))))[name = string("layers_6_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280838400))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293433728))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293421376))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293439936))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297638464))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297634304))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297640576))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165504))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298164928))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165824))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302364352))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302360192))))[name = string("layers_7_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302366464))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314961792))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314949440))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(1)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_280 = const()[name = string("op_280"), val = int32(0)];
            bool var_282_exclusive_0 = const()[name = string("op_282_exclusive_0"), val = bool(false)];
            bool var_282_reverse_0 = const()[name = string("op_282_reverse_0"), val = bool(false)];
            tensor<fp16, [1]> var_282_cast_fp16 = cumsum(axis = var_280, exclusive = var_282_exclusive_0, reverse = var_282_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_282_cast_fp16")];
            fp16 var_284_promoted_to_fp16 = const()[name = string("op_284_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> position_offsets_cast_fp16 = sub(x = var_282_cast_fp16, y = var_284_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1]> var_287_cast_fp16 = expand_dims(axes = var_287_axes_0, x = position_offsets_cast_fp16)[name = string("op_287_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_7")];
            tensor<fp16, [1, 1]> position_ids_1_cast_fp16 = add(x = var_287_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 1]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_6")];
            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 1]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 1]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314968000)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_2_y_0 = const()[name = string("greater_equal_2_y_0"), val = int32(0)];
            tensor<bool, [1, 1]> greater_equal_2 = greater_equal(x = select_0, y = greater_equal_2_y_0)[name = string("greater_equal_2")];
            int32 slice_by_index_2 = const()[name = string("slice_by_index_2"), val = int32(32768)];
            tensor<int32, [1, 1]> add_2 = add(x = select_0, y = slice_by_index_2)[name = string("add_2")];
            tensor<int32, [1, 1]> select_2 = select(a = select_0, b = add_2, cond = greater_equal_2)[name = string("select_2")];
            int32 cos_1_cast_fp16_axis_1 = const()[name = string("cos_1_cast_fp16_axis_1"), val = int32(0)];
            tensor<fp16, [1, 1, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_1, batch_dims = cos_1_batch_dims_0, indices = select_2, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323356672)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_1 = const()[name = string("sin_1_cast_fp16_axis_1"), val = int32(0)];
            tensor<fp16, [1, 1, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_1, batch_dims = sin_1_batch_dims_0, indices = select_2, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_306_perm_0 = const()[name = string("op_306_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_308_axes_0 = const()[name = string("op_308_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_306_cast_fp16 = transpose(perm = var_306_perm_0, x = cos_1_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 1, 128, 1]> var_308_cast_fp16 = expand_dims(axes = var_308_axes_0, x = var_306_cast_fp16)[name = string("op_308_cast_fp16")];
            tensor<int32, [3]> var_313_perm_0 = const()[name = string("op_313_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_313_cast_fp16 = transpose(perm = var_313_perm_0, x = sin_1_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 1, 128, 1]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = var_313_cast_fp16)[name = string("op_315_cast_fp16")];
            tensor<int32, [1]> var_334_axes_0 = const()[name = string("op_334_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 1, 1]> var_334 = expand_dims(axes = var_334_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_334")];
            tensor<int32, [1, 1, 2048]> var_327 = const()[name = string("op_327"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331745344)))];
            tensor<bool, [1, 1, 2048]> var_335 = greater(x = var_327, y = var_334)[name = string("op_335")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_342_axes_0 = const()[name = string("op_342_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 2048]> var_335_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_335)[name = string("cast_5")];
            tensor<fp16, [1, 1, 1, 2048]> var_342_cast_fp16 = expand_dims(axes = var_342_axes_0, x = var_335_to_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_346_promoted_to_fp16 = const()[name = string("op_346_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 1]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_342_cast_fp16)[name = string("transpose_57")];
            tensor<bool, [1, 1, 2048, 1]> var_347_cast_fp16 = equal(x = mask_cast_fp16, y = var_346_promoted_to_fp16)[name = string("op_347_cast_fp16")];
            fp16 var_348_to_fp16 = const()[name = string("op_348_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 1]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_348_to_fp16, cond = var_347_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_4")];
            tensor<fp16, [1, 2048, 1, 1]> var_358_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_358_cast_fp16")];
            int32 var_356 = const()[name = string("op_356"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_1_cast_fp16 = concat(axis = var_356, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_358_cast_fp16))[name = string("doubled_1_cast_fp16")];
            tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_1_gamma_0_to_fp16 = const()[name = string("out_1_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331753600)))];
            fp16 var_368_to_fp16 = const()[name = string("op_368_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_368_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_379_split_sizes_0 = const()[name = string("op_379_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_379_axis_0 = const()[name = string("op_379_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_379_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_379_cast_fp16_1 = split(axis = var_379_axis_0, split_sizes = var_379_split_sizes_0, x = out_1_cast_fp16)[name = string("op_379_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331761856)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_379_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_436_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_436_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_443_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_443_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_447_cast_fp16 = mul(x = x_1_cast_fp16, y = var_308_cast_fp16)[name = string("op_447_cast_fp16")];
            tensor<int32, [2]> var_448_split_sizes_0 = const()[name = string("op_448_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_448_axis_0 = const()[name = string("op_448_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_448_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_448_cast_fp16_1 = split(axis = var_448_axis_0, split_sizes = var_448_split_sizes_0, x = x_1_cast_fp16)[name = string("op_448_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_450_cast_fp16 = mul(x = var_448_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_450_cast_fp16")];
            int32 var_452 = const()[name = string("op_452"), val = int32(-2)];
            bool var_453_interleave_0 = const()[name = string("op_453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_453_cast_fp16 = concat(axis = var_452, interleave = var_453_interleave_0, values = (var_450_cast_fp16, var_448_cast_fp16_0))[name = string("op_453_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_454_cast_fp16 = mul(x = var_453_cast_fp16, y = var_315_cast_fp16)[name = string("op_454_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_3_cast_fp16 = add(x = var_447_cast_fp16, y = var_454_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_460_cast_fp16 = mul(x = var_436_cast_fp16, y = var_308_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [2]> var_461_split_sizes_0 = const()[name = string("op_461_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_461_axis_0 = const()[name = string("op_461_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_461_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_461_cast_fp16_1 = split(axis = var_461_axis_0, split_sizes = var_461_split_sizes_0, x = var_436_cast_fp16)[name = string("op_461_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_463_cast_fp16 = mul(x = var_461_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_463_cast_fp16")];
            int32 var_465 = const()[name = string("op_465"), val = int32(-2)];
            bool var_466_interleave_0 = const()[name = string("op_466_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_466_cast_fp16 = concat(axis = var_465, interleave = var_466_interleave_0, values = (var_463_cast_fp16, var_461_cast_fp16_0))[name = string("op_466_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_467_cast_fp16 = mul(x = var_466_cast_fp16, y = var_315_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_5_cast_fp16 = add(x = var_460_cast_fp16, y = var_467_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_18_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_18 = read_state(input = key_cache)[name = string("coreml_update_state_18")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_443_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_19_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_19 = read_state(input = value_cache)[name = string("coreml_update_state_19")];
            tensor<int32, [4]> var_537_begin_0 = const()[name = string("op_537_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_537_end_0 = const()[name = string("op_537_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_537_end_mask_0 = const()[name = string("op_537_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_537_cast_fp16 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = coreml_update_state_18)[name = string("op_537_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_540_axis_0 = const()[name = string("op_540_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_1 = split(axis = var_540_axis_0, split_sizes = tile_0, x = var_537_cast_fp16)[name = string("op_540_cast_fp16")];
            tensor<int32, [4]> var_547_begin_0 = const()[name = string("op_547_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_547_end_0 = const()[name = string("op_547_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_547_end_mask_0 = const()[name = string("op_547_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_547_cast_fp16 = slice_by_index(begin = var_547_begin_0, end = var_547_end_0, end_mask = var_547_end_mask_0, x = coreml_update_state_19)[name = string("op_547_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_550_axis_0 = const()[name = string("op_550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_1 = split(axis = var_550_axis_0, split_sizes = tile_1, x = var_547_cast_fp16)[name = string("op_550_cast_fp16")];
            tensor<int32, [2]> var_553_split_sizes_0 = const()[name = string("op_553_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_553_axis_0 = const()[name = string("op_553_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_553_0, tensor<fp16, [1, 8, 128, 1]> var_553_1 = split(axis = var_553_axis_0, split_sizes = var_553_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_553")];
            bool attn_weights_1_transpose_x_0 = const()[name = string("attn_weights_1_transpose_x_0"), val = bool(false)];
            bool attn_weights_1_transpose_y_0 = const()[name = string("attn_weights_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_540_cast_fp16_0, y = var_553_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_556_to_fp16 = const()[name = string("op_556_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_556_to_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_5_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
            int32 var_560 = const()[name = string("op_560"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_7_cast_fp16 = softmax(axis = var_560, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_566_transpose_x_1 = const()[name = string("op_566_transpose_x_1"), val = bool(true)];
            bool var_566_transpose_y_1 = const()[name = string("op_566_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_566_cast_fp16 = matmul(transpose_x = var_566_transpose_x_1, transpose_y = var_566_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_550_cast_fp16_0)[name = string("op_566_cast_fp16")];
            bool attn_weights_9_transpose_x_0 = const()[name = string("attn_weights_9_transpose_x_0"), val = bool(false)];
            bool attn_weights_9_transpose_y_0 = const()[name = string("attn_weights_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_9_cast_fp16 = matmul(transpose_x = attn_weights_9_transpose_x_0, transpose_y = attn_weights_9_transpose_y_0, x = var_540_cast_fp16_1, y = var_553_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_568_to_fp16 = const()[name = string("op_568_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_11_cast_fp16 = mul(x = attn_weights_9_cast_fp16, y = var_568_to_fp16)[name = string("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_13_cast_fp16 = add(x = attn_weights_11_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_13_cast_fp16")];
            int32 var_572 = const()[name = string("op_572"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_15_cast_fp16 = softmax(axis = var_572, x = attn_weights_13_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(true)];
            bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_15_cast_fp16, y = var_550_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_580 = const()[name = string("op_580"), val = int32(1)];
            bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_3_cast_fp16 = concat(axis = var_580, interleave = attn_output_3_interleave_0, values = (var_566_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_584_perm_0 = const()[name = string("op_584_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_584_cast_fp16 = transpose(perm = var_584_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_584_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332810496)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_617_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_617_cast_fp16")];
            int32 var_615 = const()[name = string("op_615"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_5_cast_fp16 = concat(axis = var_615, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_617_cast_fp16))[name = string("doubled_5_cast_fp16")];
            tensor<int32, [1]> out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_3_gamma_0_to_fp16 = const()[name = string("out_3_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341199168)))];
            fp16 var_627_to_fp16 = const()[name = string("op_627_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_627_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_638_split_sizes_0 = const()[name = string("op_638_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_638_axis_0 = const()[name = string("op_638_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_638_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_638_cast_fp16_1 = split(axis = var_638_axis_0, split_sizes = var_638_split_sizes_0, x = out_3_cast_fp16)[name = string("op_638_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_655_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_655_cast_fp16")];
            tensor<int32, [2]> var_661_strides_0 = const()[name = string("op_661_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_661_pad_type_0 = const()[name = string("op_661_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_661_pad_0 = const()[name = string("op_661_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_661_dilations_0 = const()[name = string("op_661_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_661_groups_0 = const()[name = string("op_661_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_661_cast_fp16 = conv(dilations = var_661_dilations_0, groups = var_661_groups_0, pad = var_661_pad_0, pad_type = var_661_pad_type_0, strides = var_661_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("op_661_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_9_cast_fp16 = mul(x = var_655_cast_fp16, y = var_661_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_679_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_679_cast_fp16")];
            int32 var_677 = const()[name = string("op_677"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_9_cast_fp16 = concat(axis = var_677, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_679_cast_fp16))[name = string("doubled_9_cast_fp16")];
            tensor<int32, [1]> out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_5_gamma_0_to_fp16 = const()[name = string("out_5_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341207424)))];
            fp16 var_689_to_fp16 = const()[name = string("op_689_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_689_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_700_split_sizes_0 = const()[name = string("op_700_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_700_axis_0 = const()[name = string("op_700_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_700_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_700_cast_fp16_1 = split(axis = var_700_axis_0, split_sizes = var_700_split_sizes_0, x = out_5_cast_fp16)[name = string("op_700_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341215680)))];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = var_700_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_757_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_757_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_764_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_764_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_768_cast_fp16 = mul(x = x_11_cast_fp16, y = var_308_cast_fp16)[name = string("op_768_cast_fp16")];
            tensor<int32, [2]> var_769_split_sizes_0 = const()[name = string("op_769_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_769_axis_0 = const()[name = string("op_769_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_769_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_769_cast_fp16_1 = split(axis = var_769_axis_0, split_sizes = var_769_split_sizes_0, x = x_11_cast_fp16)[name = string("op_769_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_771_cast_fp16 = mul(x = var_769_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_771_cast_fp16")];
            int32 var_773 = const()[name = string("op_773"), val = int32(-2)];
            bool var_774_interleave_0 = const()[name = string("op_774_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_774_cast_fp16 = concat(axis = var_773, interleave = var_774_interleave_0, values = (var_771_cast_fp16, var_769_cast_fp16_0))[name = string("op_774_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_775_cast_fp16 = mul(x = var_774_cast_fp16, y = var_315_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_9_cast_fp16 = add(x = var_768_cast_fp16, y = var_775_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_781_cast_fp16 = mul(x = var_757_cast_fp16, y = var_308_cast_fp16)[name = string("op_781_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_782_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = var_757_cast_fp16)[name = string("op_782_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_784_cast_fp16 = mul(x = var_782_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_784_cast_fp16")];
            int32 var_786 = const()[name = string("op_786"), val = int32(-2)];
            bool var_787_interleave_0 = const()[name = string("op_787_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_787_cast_fp16 = concat(axis = var_786, interleave = var_787_interleave_0, values = (var_784_cast_fp16, var_782_cast_fp16_0))[name = string("op_787_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_788_cast_fp16 = mul(x = var_787_cast_fp16, y = var_315_cast_fp16)[name = string("op_788_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_15_cast_fp16 = add(x = var_781_cast_fp16, y = var_788_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_18)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_20_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_20 = read_state(input = key_cache)[name = string("coreml_update_state_20")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_764_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_19)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_21_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_21 = read_state(input = value_cache)[name = string("coreml_update_state_21")];
            tensor<int32, [4]> var_858_begin_0 = const()[name = string("op_858_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_858_end_0 = const()[name = string("op_858_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_858_end_mask_0 = const()[name = string("op_858_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_858_cast_fp16 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = coreml_update_state_20)[name = string("op_858_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_861_axis_0 = const()[name = string("op_861_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_1 = split(axis = var_861_axis_0, split_sizes = tile_2, x = var_858_cast_fp16)[name = string("op_861_cast_fp16")];
            tensor<int32, [4]> var_868_begin_0 = const()[name = string("op_868_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_868_end_0 = const()[name = string("op_868_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_868_end_mask_0 = const()[name = string("op_868_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_868_cast_fp16 = slice_by_index(begin = var_868_begin_0, end = var_868_end_0, end_mask = var_868_end_mask_0, x = coreml_update_state_21)[name = string("op_868_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_871_axis_0 = const()[name = string("op_871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_1 = split(axis = var_871_axis_0, split_sizes = tile_3, x = var_868_cast_fp16)[name = string("op_871_cast_fp16")];
            tensor<int32, [2]> var_874_split_sizes_0 = const()[name = string("op_874_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_874_axis_0 = const()[name = string("op_874_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_874_0, tensor<fp16, [1, 8, 128, 1]> var_874_1 = split(axis = var_874_axis_0, split_sizes = var_874_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_874")];
            bool attn_weights_17_transpose_x_0 = const()[name = string("attn_weights_17_transpose_x_0"), val = bool(false)];
            bool attn_weights_17_transpose_y_0 = const()[name = string("attn_weights_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_17_cast_fp16 = matmul(transpose_x = attn_weights_17_transpose_x_0, transpose_y = attn_weights_17_transpose_y_0, x = var_861_cast_fp16_0, y = var_874_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_877_to_fp16 = const()[name = string("op_877_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_19_cast_fp16 = mul(x = attn_weights_17_cast_fp16, y = var_877_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            int32 var_881 = const()[name = string("op_881"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_23_cast_fp16 = softmax(axis = var_881, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_887_transpose_x_1 = const()[name = string("op_887_transpose_x_1"), val = bool(true)];
            bool var_887_transpose_y_1 = const()[name = string("op_887_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_887_cast_fp16 = matmul(transpose_x = var_887_transpose_x_1, transpose_y = var_887_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_871_cast_fp16_0)[name = string("op_887_cast_fp16")];
            bool attn_weights_25_transpose_x_0 = const()[name = string("attn_weights_25_transpose_x_0"), val = bool(false)];
            bool attn_weights_25_transpose_y_0 = const()[name = string("attn_weights_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_0, transpose_y = attn_weights_25_transpose_y_0, x = var_861_cast_fp16_1, y = var_874_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_889_to_fp16 = const()[name = string("op_889_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_889_to_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_29_cast_fp16 = add(x = attn_weights_27_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
            int32 var_893 = const()[name = string("op_893"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_31_cast_fp16 = softmax(axis = var_893, x = attn_weights_29_cast_fp16)[name = string("attn_weights_31_cast_fp16")];
            bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(true)];
            bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_31_cast_fp16, y = var_871_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_901 = const()[name = string("op_901"), val = int32(1)];
            bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_11_cast_fp16 = concat(axis = var_901, interleave = attn_output_11_interleave_0, values = (var_887_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_905_perm_0 = const()[name = string("op_905_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_905_cast_fp16 = transpose(perm = var_905_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_905_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_938_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_938_cast_fp16")];
            int32 var_936 = const()[name = string("op_936"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_13_cast_fp16 = concat(axis = var_936, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_938_cast_fp16))[name = string("doubled_13_cast_fp16")];
            tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_7_gamma_0_to_fp16 = const()[name = string("out_7_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342264320)))];
            fp16 var_948_to_fp16 = const()[name = string("op_948_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_948_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_959_split_sizes_0 = const()[name = string("op_959_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_959_axis_0 = const()[name = string("op_959_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_959_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_959_cast_fp16_1 = split(axis = var_959_axis_0, split_sizes = var_959_split_sizes_0, x = out_7_cast_fp16)[name = string("op_959_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_959_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_976_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_976_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342272576)))];
            tensor<int32, [2]> var_982_strides_0 = const()[name = string("op_982_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_982_pad_type_0 = const()[name = string("op_982_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_982_pad_0 = const()[name = string("op_982_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_982_dilations_0 = const()[name = string("op_982_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_982_groups_0 = const()[name = string("op_982_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_982_cast_fp16 = conv(dilations = var_982_dilations_0, groups = var_982_groups_0, pad = var_982_pad_0, pad_type = var_982_pad_type_0, strides = var_982_strides_0, weight = layers_1_mlp_up_proj_weight_to_fp16, x = var_959_cast_fp16_0)[name = string("op_982_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_19_cast_fp16 = mul(x = var_976_cast_fp16, y = var_982_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1000_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1000_cast_fp16")];
            int32 var_998 = const()[name = string("op_998"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_17_cast_fp16 = concat(axis = var_998, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1000_cast_fp16))[name = string("doubled_17_cast_fp16")];
            tensor<int32, [1]> out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_9_gamma_0_to_fp16 = const()[name = string("out_9_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367438464)))];
            fp16 var_1010_to_fp16 = const()[name = string("op_1010_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1010_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1021_split_sizes_0 = const()[name = string("op_1021_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1021_axis_0 = const()[name = string("op_1021_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1021_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1021_cast_fp16_1 = split(axis = var_1021_axis_0, split_sizes = var_1021_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1021_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367446720)))];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = var_1021_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1078_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1078_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1085_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1085_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1089_cast_fp16 = mul(x = x_21_cast_fp16, y = var_308_cast_fp16)[name = string("op_1089_cast_fp16")];
            tensor<int32, [2]> var_1090_split_sizes_0 = const()[name = string("op_1090_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1090_axis_0 = const()[name = string("op_1090_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1090_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1090_cast_fp16_1 = split(axis = var_1090_axis_0, split_sizes = var_1090_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1090_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_1092_cast_fp16 = mul(x = var_1090_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            int32 var_1094 = const()[name = string("op_1094"), val = int32(-2)];
            bool var_1095_interleave_0 = const()[name = string("op_1095_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1095_cast_fp16 = concat(axis = var_1094, interleave = var_1095_interleave_0, values = (var_1092_cast_fp16, var_1090_cast_fp16_0))[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1096_cast_fp16 = mul(x = var_1095_cast_fp16, y = var_315_cast_fp16)[name = string("op_1096_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_15_cast_fp16 = add(x = var_1089_cast_fp16, y = var_1096_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1102_cast_fp16 = mul(x = var_1078_cast_fp16, y = var_308_cast_fp16)[name = string("op_1102_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1103_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = var_1078_cast_fp16)[name = string("op_1103_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_1105_cast_fp16 = mul(x = var_1103_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1105_cast_fp16")];
            int32 var_1107 = const()[name = string("op_1107"), val = int32(-2)];
            bool var_1108_interleave_0 = const()[name = string("op_1108_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1108_cast_fp16 = concat(axis = var_1107, interleave = var_1108_interleave_0, values = (var_1105_cast_fp16, var_1103_cast_fp16_0))[name = string("op_1108_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1109_cast_fp16 = mul(x = var_1108_cast_fp16, y = var_315_cast_fp16)[name = string("op_1109_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_25_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1109_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_20)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_22_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_22 = read_state(input = key_cache)[name = string("coreml_update_state_22")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1085_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_21)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_23_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_23 = read_state(input = value_cache)[name = string("coreml_update_state_23")];
            tensor<int32, [4]> var_1179_begin_0 = const()[name = string("op_1179_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1179_end_0 = const()[name = string("op_1179_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1179_end_mask_0 = const()[name = string("op_1179_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1179_cast_fp16 = slice_by_index(begin = var_1179_begin_0, end = var_1179_end_0, end_mask = var_1179_end_mask_0, x = coreml_update_state_22)[name = string("op_1179_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1182_axis_0 = const()[name = string("op_1182_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_1 = split(axis = var_1182_axis_0, split_sizes = tile_4, x = var_1179_cast_fp16)[name = string("op_1182_cast_fp16")];
            tensor<int32, [4]> var_1189_begin_0 = const()[name = string("op_1189_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1189_end_0 = const()[name = string("op_1189_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1189_end_mask_0 = const()[name = string("op_1189_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1189_cast_fp16 = slice_by_index(begin = var_1189_begin_0, end = var_1189_end_0, end_mask = var_1189_end_mask_0, x = coreml_update_state_23)[name = string("op_1189_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1192_axis_0 = const()[name = string("op_1192_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_1 = split(axis = var_1192_axis_0, split_sizes = tile_5, x = var_1189_cast_fp16)[name = string("op_1192_cast_fp16")];
            tensor<int32, [2]> var_1195_split_sizes_0 = const()[name = string("op_1195_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1195_axis_0 = const()[name = string("op_1195_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1195_0, tensor<fp16, [1, 8, 128, 1]> var_1195_1 = split(axis = var_1195_axis_0, split_sizes = var_1195_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1195")];
            bool attn_weights_33_transpose_x_0 = const()[name = string("attn_weights_33_transpose_x_0"), val = bool(false)];
            bool attn_weights_33_transpose_y_0 = const()[name = string("attn_weights_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_33_cast_fp16 = matmul(transpose_x = attn_weights_33_transpose_x_0, transpose_y = attn_weights_33_transpose_y_0, x = var_1182_cast_fp16_0, y = var_1195_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1198_to_fp16 = const()[name = string("op_1198_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_35_cast_fp16 = mul(x = attn_weights_33_cast_fp16, y = var_1198_to_fp16)[name = string("attn_weights_35_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_37_cast_fp16 = add(x = attn_weights_35_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_37_cast_fp16")];
            int32 var_1202 = const()[name = string("op_1202"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_39_cast_fp16 = softmax(axis = var_1202, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1208_transpose_x_1 = const()[name = string("op_1208_transpose_x_1"), val = bool(true)];
            bool var_1208_transpose_y_1 = const()[name = string("op_1208_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1208_cast_fp16 = matmul(transpose_x = var_1208_transpose_x_1, transpose_y = var_1208_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1192_cast_fp16_0)[name = string("op_1208_cast_fp16")];
            bool attn_weights_41_transpose_x_0 = const()[name = string("attn_weights_41_transpose_x_0"), val = bool(false)];
            bool attn_weights_41_transpose_y_0 = const()[name = string("attn_weights_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_41_cast_fp16 = matmul(transpose_x = attn_weights_41_transpose_x_0, transpose_y = attn_weights_41_transpose_y_0, x = var_1182_cast_fp16_1, y = var_1195_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1210_to_fp16 = const()[name = string("op_1210_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_43_cast_fp16 = mul(x = attn_weights_41_cast_fp16, y = var_1210_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            int32 var_1214 = const()[name = string("op_1214"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_47_cast_fp16 = softmax(axis = var_1214, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")];
            bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(true)];
            bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_47_cast_fp16, y = var_1192_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1222 = const()[name = string("op_1222"), val = int32(1)];
            bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_19_cast_fp16 = concat(axis = var_1222, interleave = attn_output_19_interleave_0, values = (var_1208_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1226_perm_0 = const()[name = string("op_1226_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_1226_cast_fp16 = transpose(perm = var_1226_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1226_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1259_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            int32 var_1257 = const()[name = string("op_1257"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_21_cast_fp16 = concat(axis = var_1257, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1259_cast_fp16))[name = string("doubled_21_cast_fp16")];
            tensor<int32, [1]> out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_11_gamma_0_to_fp16 = const()[name = string("out_11_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368495360)))];
            fp16 var_1269_to_fp16 = const()[name = string("op_1269_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1269_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1280_split_sizes_0 = const()[name = string("op_1280_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1280_axis_0 = const()[name = string("op_1280_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1280_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1280_cast_fp16_1 = split(axis = var_1280_axis_0, split_sizes = var_1280_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1280_cast_fp16")];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1297_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1297_cast_fp16")];
            tensor<int32, [2]> var_1303_strides_0 = const()[name = string("op_1303_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1303_pad_type_0 = const()[name = string("op_1303_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1303_pad_0 = const()[name = string("op_1303_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1303_dilations_0 = const()[name = string("op_1303_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1303_groups_0 = const()[name = string("op_1303_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1303_cast_fp16 = conv(dilations = var_1303_dilations_0, groups = var_1303_groups_0, pad = var_1303_pad_0, pad_type = var_1303_pad_type_0, strides = var_1303_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("op_1303_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_29_cast_fp16 = mul(x = var_1297_cast_fp16, y = var_1303_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1321_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1321_cast_fp16")];
            int32 var_1319 = const()[name = string("op_1319"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_25_cast_fp16 = concat(axis = var_1319, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1321_cast_fp16))[name = string("doubled_25_cast_fp16")];
            tensor<int32, [1]> out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_13_gamma_0_to_fp16 = const()[name = string("out_13_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368503616)))];
            fp16 var_1331_to_fp16 = const()[name = string("op_1331_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1331_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1342_split_sizes_0 = const()[name = string("op_1342_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1342_axis_0 = const()[name = string("op_1342_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1342_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1342_cast_fp16_1 = split(axis = var_1342_axis_0, split_sizes = var_1342_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1342_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368511872)))];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = var_1342_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1399_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1406_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1406_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1410_cast_fp16 = mul(x = x_31_cast_fp16, y = var_308_cast_fp16)[name = string("op_1410_cast_fp16")];
            tensor<int32, [2]> var_1411_split_sizes_0 = const()[name = string("op_1411_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1411_axis_0 = const()[name = string("op_1411_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1411_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1411_cast_fp16_1 = split(axis = var_1411_axis_0, split_sizes = var_1411_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1411_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_1413_cast_fp16 = mul(x = var_1411_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1413_cast_fp16")];
            int32 var_1415 = const()[name = string("op_1415"), val = int32(-2)];
            bool var_1416_interleave_0 = const()[name = string("op_1416_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1416_cast_fp16 = concat(axis = var_1415, interleave = var_1416_interleave_0, values = (var_1413_cast_fp16, var_1411_cast_fp16_0))[name = string("op_1416_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1417_cast_fp16 = mul(x = var_1416_cast_fp16, y = var_315_cast_fp16)[name = string("op_1417_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_21_cast_fp16 = add(x = var_1410_cast_fp16, y = var_1417_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1423_cast_fp16 = mul(x = var_1399_cast_fp16, y = var_308_cast_fp16)[name = string("op_1423_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1424_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = var_1399_cast_fp16)[name = string("op_1424_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_1426_cast_fp16 = mul(x = var_1424_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1426_cast_fp16")];
            int32 var_1428 = const()[name = string("op_1428"), val = int32(-2)];
            bool var_1429_interleave_0 = const()[name = string("op_1429_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1429_cast_fp16 = concat(axis = var_1428, interleave = var_1429_interleave_0, values = (var_1426_cast_fp16, var_1424_cast_fp16_0))[name = string("op_1429_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1430_cast_fp16 = mul(x = var_1429_cast_fp16, y = var_315_cast_fp16)[name = string("op_1430_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_35_cast_fp16 = add(x = var_1423_cast_fp16, y = var_1430_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_22)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_24_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_24 = read_state(input = key_cache)[name = string("coreml_update_state_24")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1406_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_23)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_25_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_25 = read_state(input = value_cache)[name = string("coreml_update_state_25")];
            tensor<int32, [4]> var_1500_begin_0 = const()[name = string("op_1500_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1500_end_0 = const()[name = string("op_1500_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1500_end_mask_0 = const()[name = string("op_1500_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1500_cast_fp16 = slice_by_index(begin = var_1500_begin_0, end = var_1500_end_0, end_mask = var_1500_end_mask_0, x = coreml_update_state_24)[name = string("op_1500_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1503_axis_0 = const()[name = string("op_1503_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_1 = split(axis = var_1503_axis_0, split_sizes = tile_6, x = var_1500_cast_fp16)[name = string("op_1503_cast_fp16")];
            tensor<int32, [4]> var_1510_begin_0 = const()[name = string("op_1510_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1510_end_0 = const()[name = string("op_1510_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1510_end_mask_0 = const()[name = string("op_1510_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1510_cast_fp16 = slice_by_index(begin = var_1510_begin_0, end = var_1510_end_0, end_mask = var_1510_end_mask_0, x = coreml_update_state_25)[name = string("op_1510_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1513_axis_0 = const()[name = string("op_1513_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_1 = split(axis = var_1513_axis_0, split_sizes = tile_7, x = var_1510_cast_fp16)[name = string("op_1513_cast_fp16")];
            tensor<int32, [2]> var_1516_split_sizes_0 = const()[name = string("op_1516_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1516_axis_0 = const()[name = string("op_1516_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1516_0, tensor<fp16, [1, 8, 128, 1]> var_1516_1 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1516")];
            bool attn_weights_49_transpose_x_0 = const()[name = string("attn_weights_49_transpose_x_0"), val = bool(false)];
            bool attn_weights_49_transpose_y_0 = const()[name = string("attn_weights_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_0, transpose_y = attn_weights_49_transpose_y_0, x = var_1503_cast_fp16_0, y = var_1516_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_1519_to_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_53_cast_fp16 = add(x = attn_weights_51_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_53_cast_fp16")];
            int32 var_1523 = const()[name = string("op_1523"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_55_cast_fp16 = softmax(axis = var_1523, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1529_transpose_x_1 = const()[name = string("op_1529_transpose_x_1"), val = bool(true)];
            bool var_1529_transpose_y_1 = const()[name = string("op_1529_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1529_cast_fp16 = matmul(transpose_x = var_1529_transpose_x_1, transpose_y = var_1529_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1513_cast_fp16_0)[name = string("op_1529_cast_fp16")];
            bool attn_weights_57_transpose_x_0 = const()[name = string("attn_weights_57_transpose_x_0"), val = bool(false)];
            bool attn_weights_57_transpose_y_0 = const()[name = string("attn_weights_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_57_cast_fp16 = matmul(transpose_x = attn_weights_57_transpose_x_0, transpose_y = attn_weights_57_transpose_y_0, x = var_1503_cast_fp16_1, y = var_1516_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1531_to_fp16 = const()[name = string("op_1531_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_59_cast_fp16 = mul(x = attn_weights_57_cast_fp16, y = var_1531_to_fp16)[name = string("attn_weights_59_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_61_cast_fp16 = add(x = attn_weights_59_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_61_cast_fp16")];
            int32 var_1535 = const()[name = string("op_1535"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_63_cast_fp16 = softmax(axis = var_1535, x = attn_weights_61_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(true)];
            bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_63_cast_fp16, y = var_1513_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1543 = const()[name = string("op_1543"), val = int32(1)];
            bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_27_cast_fp16 = concat(axis = var_1543, interleave = attn_output_27_interleave_0, values = (var_1529_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1547_perm_0 = const()[name = string("op_1547_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_1547_cast_fp16 = transpose(perm = var_1547_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1547_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369560512)))];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1580_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1580_cast_fp16")];
            int32 var_1578 = const()[name = string("op_1578"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_29_cast_fp16 = concat(axis = var_1578, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1580_cast_fp16))[name = string("doubled_29_cast_fp16")];
            tensor<int32, [1]> out_15_axes_0 = const()[name = string("out_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_15_gamma_0_to_fp16 = const()[name = string("out_15_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377949184)))];
            fp16 var_1590_to_fp16 = const()[name = string("op_1590_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_1590_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1601_split_sizes_0 = const()[name = string("op_1601_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1601_axis_0 = const()[name = string("op_1601_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1601_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1601_cast_fp16_1 = split(axis = var_1601_axis_0, split_sizes = var_1601_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1601_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1618_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1618_cast_fp16")];
            tensor<int32, [2]> var_1624_strides_0 = const()[name = string("op_1624_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1624_pad_type_0 = const()[name = string("op_1624_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1624_pad_0 = const()[name = string("op_1624_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1624_dilations_0 = const()[name = string("op_1624_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1624_groups_0 = const()[name = string("op_1624_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1624_cast_fp16 = conv(dilations = var_1624_dilations_0, groups = var_1624_groups_0, pad = var_1624_pad_0, pad_type = var_1624_pad_type_0, strides = var_1624_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("op_1624_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_39_cast_fp16 = mul(x = var_1618_cast_fp16, y = var_1624_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1642_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1642_cast_fp16")];
            int32 var_1640 = const()[name = string("op_1640"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_33_cast_fp16 = concat(axis = var_1640, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1642_cast_fp16))[name = string("doubled_33_cast_fp16")];
            tensor<int32, [1]> out_17_axes_0 = const()[name = string("out_17_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_17_gamma_0_to_fp16 = const()[name = string("out_17_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377957440)))];
            fp16 var_1652_to_fp16 = const()[name = string("op_1652_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1652_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1663_split_sizes_0 = const()[name = string("op_1663_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1663_axis_0 = const()[name = string("op_1663_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1663_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1663_cast_fp16_1 = split(axis = var_1663_axis_0, split_sizes = var_1663_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1663_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377965696)))];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = var_1663_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1720_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1720_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1727_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1731_cast_fp16 = mul(x = x_41_cast_fp16, y = var_308_cast_fp16)[name = string("op_1731_cast_fp16")];
            tensor<int32, [2]> var_1732_split_sizes_0 = const()[name = string("op_1732_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1732_axis_0 = const()[name = string("op_1732_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1732_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1732_cast_fp16_1 = split(axis = var_1732_axis_0, split_sizes = var_1732_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1732_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_1734_cast_fp16 = mul(x = var_1732_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1734_cast_fp16")];
            int32 var_1736 = const()[name = string("op_1736"), val = int32(-2)];
            bool var_1737_interleave_0 = const()[name = string("op_1737_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1737_cast_fp16 = concat(axis = var_1736, interleave = var_1737_interleave_0, values = (var_1734_cast_fp16, var_1732_cast_fp16_0))[name = string("op_1737_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1738_cast_fp16 = mul(x = var_1737_cast_fp16, y = var_315_cast_fp16)[name = string("op_1738_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_27_cast_fp16 = add(x = var_1731_cast_fp16, y = var_1738_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1744_cast_fp16 = mul(x = var_1720_cast_fp16, y = var_308_cast_fp16)[name = string("op_1744_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1745_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = var_1720_cast_fp16)[name = string("op_1745_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_1747_cast_fp16 = mul(x = var_1745_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1747_cast_fp16")];
            int32 var_1749 = const()[name = string("op_1749"), val = int32(-2)];
            bool var_1750_interleave_0 = const()[name = string("op_1750_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1750_cast_fp16 = concat(axis = var_1749, interleave = var_1750_interleave_0, values = (var_1747_cast_fp16, var_1745_cast_fp16_0))[name = string("op_1750_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1751_cast_fp16 = mul(x = var_1750_cast_fp16, y = var_315_cast_fp16)[name = string("op_1751_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_45_cast_fp16 = add(x = var_1744_cast_fp16, y = var_1751_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_24)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_26_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_26 = read_state(input = key_cache)[name = string("coreml_update_state_26")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1727_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_25)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_27_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_27 = read_state(input = value_cache)[name = string("coreml_update_state_27")];
            tensor<int32, [4]> var_1821_begin_0 = const()[name = string("op_1821_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1821_end_0 = const()[name = string("op_1821_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1821_end_mask_0 = const()[name = string("op_1821_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1821_cast_fp16 = slice_by_index(begin = var_1821_begin_0, end = var_1821_end_0, end_mask = var_1821_end_mask_0, x = coreml_update_state_26)[name = string("op_1821_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1824_axis_0 = const()[name = string("op_1824_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_1 = split(axis = var_1824_axis_0, split_sizes = tile_8, x = var_1821_cast_fp16)[name = string("op_1824_cast_fp16")];
            tensor<int32, [4]> var_1831_begin_0 = const()[name = string("op_1831_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1831_end_0 = const()[name = string("op_1831_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1831_end_mask_0 = const()[name = string("op_1831_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1831_cast_fp16 = slice_by_index(begin = var_1831_begin_0, end = var_1831_end_0, end_mask = var_1831_end_mask_0, x = coreml_update_state_27)[name = string("op_1831_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1834_axis_0 = const()[name = string("op_1834_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_1 = split(axis = var_1834_axis_0, split_sizes = tile_9, x = var_1831_cast_fp16)[name = string("op_1834_cast_fp16")];
            tensor<int32, [2]> var_1837_split_sizes_0 = const()[name = string("op_1837_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1837_axis_0 = const()[name = string("op_1837_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1837_0, tensor<fp16, [1, 8, 128, 1]> var_1837_1 = split(axis = var_1837_axis_0, split_sizes = var_1837_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1837")];
            bool attn_weights_65_transpose_x_0 = const()[name = string("attn_weights_65_transpose_x_0"), val = bool(false)];
            bool attn_weights_65_transpose_y_0 = const()[name = string("attn_weights_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_65_cast_fp16 = matmul(transpose_x = attn_weights_65_transpose_x_0, transpose_y = attn_weights_65_transpose_y_0, x = var_1824_cast_fp16_0, y = var_1837_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1840_to_fp16 = const()[name = string("op_1840_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_67_cast_fp16 = mul(x = attn_weights_65_cast_fp16, y = var_1840_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            int32 var_1844 = const()[name = string("op_1844"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_71_cast_fp16 = softmax(axis = var_1844, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1850_transpose_x_1 = const()[name = string("op_1850_transpose_x_1"), val = bool(true)];
            bool var_1850_transpose_y_1 = const()[name = string("op_1850_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1850_cast_fp16 = matmul(transpose_x = var_1850_transpose_x_1, transpose_y = var_1850_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1834_cast_fp16_0)[name = string("op_1850_cast_fp16")];
            bool attn_weights_73_transpose_x_0 = const()[name = string("attn_weights_73_transpose_x_0"), val = bool(false)];
            bool attn_weights_73_transpose_y_0 = const()[name = string("attn_weights_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_0, transpose_y = attn_weights_73_transpose_y_0, x = var_1824_cast_fp16_1, y = var_1837_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1852_to_fp16 = const()[name = string("op_1852_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1852_to_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_77_cast_fp16 = add(x = attn_weights_75_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_77_cast_fp16")];
            int32 var_1856 = const()[name = string("op_1856"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_79_cast_fp16 = softmax(axis = var_1856, x = attn_weights_77_cast_fp16)[name = string("attn_weights_79_cast_fp16")];
            bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(true)];
            bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_79_cast_fp16, y = var_1834_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1864 = const()[name = string("op_1864"), val = int32(1)];
            bool attn_output_35_interleave_0 = const()[name = string("attn_output_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_35_cast_fp16 = concat(axis = var_1864, interleave = attn_output_35_interleave_0, values = (var_1850_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1868_perm_0 = const()[name = string("op_1868_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_1868_cast_fp16 = transpose(perm = var_1868_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1868_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1901_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1901_cast_fp16")];
            int32 var_1899 = const()[name = string("op_1899"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_37_cast_fp16 = concat(axis = var_1899, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1901_cast_fp16))[name = string("doubled_37_cast_fp16")];
            tensor<int32, [1]> out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_19_gamma_0_to_fp16 = const()[name = string("out_19_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379014336)))];
            fp16 var_1911_to_fp16 = const()[name = string("op_1911_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1911_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1922_split_sizes_0 = const()[name = string("op_1922_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1922_axis_0 = const()[name = string("op_1922_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1922_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1922_cast_fp16_1 = split(axis = var_1922_axis_0, split_sizes = var_1922_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1922_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1939_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<int32, [2]> var_1945_strides_0 = const()[name = string("op_1945_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1945_pad_type_0 = const()[name = string("op_1945_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1945_pad_0 = const()[name = string("op_1945_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1945_dilations_0 = const()[name = string("op_1945_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1945_groups_0 = const()[name = string("op_1945_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1945_cast_fp16 = conv(dilations = var_1945_dilations_0, groups = var_1945_groups_0, pad = var_1945_pad_0, pad_type = var_1945_pad_type_0, strides = var_1945_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("op_1945_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_49_cast_fp16 = mul(x = var_1939_cast_fp16, y = var_1945_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1963_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1963_cast_fp16")];
            int32 var_1961 = const()[name = string("op_1961"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_41_cast_fp16 = concat(axis = var_1961, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1963_cast_fp16))[name = string("doubled_41_cast_fp16")];
            tensor<int32, [1]> out_21_axes_0 = const()[name = string("out_21_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_21_gamma_0_to_fp16 = const()[name = string("out_21_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379022592)))];
            fp16 var_1973_to_fp16 = const()[name = string("op_1973_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1973_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_1984_split_sizes_0 = const()[name = string("op_1984_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1984_axis_0 = const()[name = string("op_1984_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1984_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1984_cast_fp16_1 = split(axis = var_1984_axis_0, split_sizes = var_1984_split_sizes_0, x = out_21_cast_fp16)[name = string("op_1984_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379030848)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_1984_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2041_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2041_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2048_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2048_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2052_cast_fp16 = mul(x = x_51_cast_fp16, y = var_308_cast_fp16)[name = string("op_2052_cast_fp16")];
            tensor<int32, [2]> var_2053_split_sizes_0 = const()[name = string("op_2053_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2053_axis_0 = const()[name = string("op_2053_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2053_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2053_cast_fp16_1 = split(axis = var_2053_axis_0, split_sizes = var_2053_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2053_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_2055_cast_fp16 = mul(x = var_2053_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2055_cast_fp16")];
            int32 var_2057 = const()[name = string("op_2057"), val = int32(-2)];
            bool var_2058_interleave_0 = const()[name = string("op_2058_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2058_cast_fp16 = concat(axis = var_2057, interleave = var_2058_interleave_0, values = (var_2055_cast_fp16, var_2053_cast_fp16_0))[name = string("op_2058_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2059_cast_fp16 = mul(x = var_2058_cast_fp16, y = var_315_cast_fp16)[name = string("op_2059_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_33_cast_fp16 = add(x = var_2052_cast_fp16, y = var_2059_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2065_cast_fp16 = mul(x = var_2041_cast_fp16, y = var_308_cast_fp16)[name = string("op_2065_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2066_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = var_2041_cast_fp16)[name = string("op_2066_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_2068_cast_fp16 = mul(x = var_2066_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2068_cast_fp16")];
            int32 var_2070 = const()[name = string("op_2070"), val = int32(-2)];
            bool var_2071_interleave_0 = const()[name = string("op_2071_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2071_cast_fp16 = concat(axis = var_2070, interleave = var_2071_interleave_0, values = (var_2068_cast_fp16, var_2066_cast_fp16_0))[name = string("op_2071_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2072_cast_fp16 = mul(x = var_2071_cast_fp16, y = var_315_cast_fp16)[name = string("op_2072_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_55_cast_fp16 = add(x = var_2065_cast_fp16, y = var_2072_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_26)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_28_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_28 = read_state(input = key_cache)[name = string("coreml_update_state_28")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2048_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_27)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_29_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_29 = read_state(input = value_cache)[name = string("coreml_update_state_29")];
            tensor<int32, [4]> var_2142_begin_0 = const()[name = string("op_2142_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2142_end_0 = const()[name = string("op_2142_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2142_end_mask_0 = const()[name = string("op_2142_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2142_cast_fp16 = slice_by_index(begin = var_2142_begin_0, end = var_2142_end_0, end_mask = var_2142_end_mask_0, x = coreml_update_state_28)[name = string("op_2142_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2145_axis_0 = const()[name = string("op_2145_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_1 = split(axis = var_2145_axis_0, split_sizes = tile_10, x = var_2142_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [4]> var_2152_begin_0 = const()[name = string("op_2152_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2152_end_0 = const()[name = string("op_2152_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2152_end_mask_0 = const()[name = string("op_2152_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2152_cast_fp16 = slice_by_index(begin = var_2152_begin_0, end = var_2152_end_0, end_mask = var_2152_end_mask_0, x = coreml_update_state_29)[name = string("op_2152_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2155_axis_0 = const()[name = string("op_2155_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_1 = split(axis = var_2155_axis_0, split_sizes = tile_11, x = var_2152_cast_fp16)[name = string("op_2155_cast_fp16")];
            tensor<int32, [2]> var_2158_split_sizes_0 = const()[name = string("op_2158_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2158_axis_0 = const()[name = string("op_2158_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2158_0, tensor<fp16, [1, 8, 128, 1]> var_2158_1 = split(axis = var_2158_axis_0, split_sizes = var_2158_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2158")];
            bool attn_weights_81_transpose_x_0 = const()[name = string("attn_weights_81_transpose_x_0"), val = bool(false)];
            bool attn_weights_81_transpose_y_0 = const()[name = string("attn_weights_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_81_cast_fp16 = matmul(transpose_x = attn_weights_81_transpose_x_0, transpose_y = attn_weights_81_transpose_y_0, x = var_2145_cast_fp16_0, y = var_2158_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2161_to_fp16 = const()[name = string("op_2161_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_83_cast_fp16 = mul(x = attn_weights_81_cast_fp16, y = var_2161_to_fp16)[name = string("attn_weights_83_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_85_cast_fp16 = add(x = attn_weights_83_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_85_cast_fp16")];
            int32 var_2165 = const()[name = string("op_2165"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_87_cast_fp16 = softmax(axis = var_2165, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2171_transpose_x_1 = const()[name = string("op_2171_transpose_x_1"), val = bool(true)];
            bool var_2171_transpose_y_1 = const()[name = string("op_2171_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2171_cast_fp16 = matmul(transpose_x = var_2171_transpose_x_1, transpose_y = var_2171_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2155_cast_fp16_0)[name = string("op_2171_cast_fp16")];
            bool attn_weights_89_transpose_x_0 = const()[name = string("attn_weights_89_transpose_x_0"), val = bool(false)];
            bool attn_weights_89_transpose_y_0 = const()[name = string("attn_weights_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_89_cast_fp16 = matmul(transpose_x = attn_weights_89_transpose_x_0, transpose_y = attn_weights_89_transpose_y_0, x = var_2145_cast_fp16_1, y = var_2158_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2173_to_fp16 = const()[name = string("op_2173_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_91_cast_fp16 = mul(x = attn_weights_89_cast_fp16, y = var_2173_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            int32 var_2177 = const()[name = string("op_2177"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_95_cast_fp16 = softmax(axis = var_2177, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")];
            bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(true)];
            bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_95_cast_fp16, y = var_2155_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2185 = const()[name = string("op_2185"), val = int32(1)];
            bool attn_output_43_interleave_0 = const()[name = string("attn_output_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_43_cast_fp16 = concat(axis = var_2185, interleave = attn_output_43_interleave_0, values = (var_2171_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2189_perm_0 = const()[name = string("op_2189_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_2189_cast_fp16 = transpose(perm = var_2189_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2189_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2222_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2222_cast_fp16")];
            int32 var_2220 = const()[name = string("op_2220"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_45_cast_fp16 = concat(axis = var_2220, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2222_cast_fp16))[name = string("doubled_45_cast_fp16")];
            tensor<int32, [1]> out_23_axes_0 = const()[name = string("out_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_23_gamma_0_to_fp16 = const()[name = string("out_23_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380079488)))];
            fp16 var_2232_to_fp16 = const()[name = string("op_2232_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_2232_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2243_split_sizes_0 = const()[name = string("op_2243_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2243_axis_0 = const()[name = string("op_2243_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2243_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2243_cast_fp16_1 = split(axis = var_2243_axis_0, split_sizes = var_2243_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2243_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2260_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<int32, [2]> var_2266_strides_0 = const()[name = string("op_2266_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2266_pad_type_0 = const()[name = string("op_2266_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2266_pad_0 = const()[name = string("op_2266_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2266_dilations_0 = const()[name = string("op_2266_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2266_groups_0 = const()[name = string("op_2266_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2266_cast_fp16 = conv(dilations = var_2266_dilations_0, groups = var_2266_groups_0, pad = var_2266_pad_0, pad_type = var_2266_pad_type_0, strides = var_2266_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("op_2266_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_59_cast_fp16 = mul(x = var_2260_cast_fp16, y = var_2266_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2284_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2284_cast_fp16")];
            int32 var_2282 = const()[name = string("op_2282"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_49_cast_fp16 = concat(axis = var_2282, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2284_cast_fp16))[name = string("doubled_49_cast_fp16")];
            tensor<int32, [1]> out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_25_gamma_0_to_fp16 = const()[name = string("out_25_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380087744)))];
            fp16 var_2294_to_fp16 = const()[name = string("op_2294_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_2294_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2305_split_sizes_0 = const()[name = string("op_2305_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2305_axis_0 = const()[name = string("op_2305_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2305_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2305_cast_fp16_1 = split(axis = var_2305_axis_0, split_sizes = var_2305_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380096000)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2305_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2362_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2362_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2369_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2369_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2373_cast_fp16 = mul(x = x_61_cast_fp16, y = var_308_cast_fp16)[name = string("op_2373_cast_fp16")];
            tensor<int32, [2]> var_2374_split_sizes_0 = const()[name = string("op_2374_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2374_axis_0 = const()[name = string("op_2374_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2374_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2374_cast_fp16_1 = split(axis = var_2374_axis_0, split_sizes = var_2374_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2376_cast_fp16")];
            int32 var_2378 = const()[name = string("op_2378"), val = int32(-2)];
            bool var_2379_interleave_0 = const()[name = string("op_2379_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2379_cast_fp16 = concat(axis = var_2378, interleave = var_2379_interleave_0, values = (var_2376_cast_fp16, var_2374_cast_fp16_0))[name = string("op_2379_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2380_cast_fp16 = mul(x = var_2379_cast_fp16, y = var_315_cast_fp16)[name = string("op_2380_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_39_cast_fp16 = add(x = var_2373_cast_fp16, y = var_2380_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2386_cast_fp16 = mul(x = var_2362_cast_fp16, y = var_308_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2387_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = var_2362_cast_fp16)[name = string("op_2387_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_2389_cast_fp16 = mul(x = var_2387_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2389_cast_fp16")];
            int32 var_2391 = const()[name = string("op_2391"), val = int32(-2)];
            bool var_2392_interleave_0 = const()[name = string("op_2392_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2392_cast_fp16 = concat(axis = var_2391, interleave = var_2392_interleave_0, values = (var_2389_cast_fp16, var_2387_cast_fp16_0))[name = string("op_2392_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2393_cast_fp16 = mul(x = var_2392_cast_fp16, y = var_315_cast_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_65_cast_fp16 = add(x = var_2386_cast_fp16, y = var_2393_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_28)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_30_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_30 = read_state(input = key_cache)[name = string("coreml_update_state_30")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2369_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_29)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_31_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_31 = read_state(input = value_cache)[name = string("coreml_update_state_31")];
            tensor<int32, [4]> var_2463_begin_0 = const()[name = string("op_2463_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2463_end_0 = const()[name = string("op_2463_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2463_end_mask_0 = const()[name = string("op_2463_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2463_cast_fp16 = slice_by_index(begin = var_2463_begin_0, end = var_2463_end_0, end_mask = var_2463_end_mask_0, x = coreml_update_state_30)[name = string("op_2463_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2466_axis_0 = const()[name = string("op_2466_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_1 = split(axis = var_2466_axis_0, split_sizes = tile_12, x = var_2463_cast_fp16)[name = string("op_2466_cast_fp16")];
            tensor<int32, [4]> var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, x = coreml_update_state_31)[name = string("op_2473_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2476_axis_0 = const()[name = string("op_2476_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_1 = split(axis = var_2476_axis_0, split_sizes = tile_13, x = var_2473_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<int32, [2]> var_2479_split_sizes_0 = const()[name = string("op_2479_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2479_axis_0 = const()[name = string("op_2479_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2479_0, tensor<fp16, [1, 8, 128, 1]> var_2479_1 = split(axis = var_2479_axis_0, split_sizes = var_2479_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2479")];
            bool attn_weights_97_transpose_x_0 = const()[name = string("attn_weights_97_transpose_x_0"), val = bool(false)];
            bool attn_weights_97_transpose_y_0 = const()[name = string("attn_weights_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_0, transpose_y = attn_weights_97_transpose_y_0, x = var_2466_cast_fp16_0, y = var_2479_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2482_to_fp16 = const()[name = string("op_2482_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_2482_to_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_101_cast_fp16 = add(x = attn_weights_99_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_101_cast_fp16")];
            int32 var_2486 = const()[name = string("op_2486"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_103_cast_fp16 = softmax(axis = var_2486, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2492_transpose_x_1 = const()[name = string("op_2492_transpose_x_1"), val = bool(true)];
            bool var_2492_transpose_y_1 = const()[name = string("op_2492_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2492_cast_fp16 = matmul(transpose_x = var_2492_transpose_x_1, transpose_y = var_2492_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2476_cast_fp16_0)[name = string("op_2492_cast_fp16")];
            bool attn_weights_105_transpose_x_0 = const()[name = string("attn_weights_105_transpose_x_0"), val = bool(false)];
            bool attn_weights_105_transpose_y_0 = const()[name = string("attn_weights_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_105_cast_fp16 = matmul(transpose_x = attn_weights_105_transpose_x_0, transpose_y = attn_weights_105_transpose_y_0, x = var_2466_cast_fp16_1, y = var_2479_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2494_to_fp16 = const()[name = string("op_2494_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_107_cast_fp16 = mul(x = attn_weights_105_cast_fp16, y = var_2494_to_fp16)[name = string("attn_weights_107_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_109_cast_fp16 = add(x = attn_weights_107_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_109_cast_fp16")];
            int32 var_2498 = const()[name = string("op_2498"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_111_cast_fp16 = softmax(axis = var_2498, x = attn_weights_109_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(true)];
            bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_111_cast_fp16, y = var_2476_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2506 = const()[name = string("op_2506"), val = int32(1)];
            bool attn_output_51_interleave_0 = const()[name = string("attn_output_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_51_cast_fp16 = concat(axis = var_2506, interleave = attn_output_51_interleave_0, values = (var_2492_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2510_perm_0 = const()[name = string("op_2510_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_2510_cast_fp16 = transpose(perm = var_2510_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2510_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2543_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2543_cast_fp16")];
            int32 var_2541 = const()[name = string("op_2541"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_53_cast_fp16 = concat(axis = var_2541, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2543_cast_fp16))[name = string("doubled_53_cast_fp16")];
            tensor<int32, [1]> out_27_axes_0 = const()[name = string("out_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_27_gamma_0_to_fp16 = const()[name = string("out_27_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381144640)))];
            fp16 var_2553_to_fp16 = const()[name = string("op_2553_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_2553_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2564_split_sizes_0 = const()[name = string("op_2564_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2564_axis_0 = const()[name = string("op_2564_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2564_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2564_cast_fp16_1 = split(axis = var_2564_axis_0, split_sizes = var_2564_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2581_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [2]> var_2587_strides_0 = const()[name = string("op_2587_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2587_pad_type_0 = const()[name = string("op_2587_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2587_pad_0 = const()[name = string("op_2587_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2587_dilations_0 = const()[name = string("op_2587_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2587_groups_0 = const()[name = string("op_2587_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2587_cast_fp16 = conv(dilations = var_2587_dilations_0, groups = var_2587_groups_0, pad = var_2587_pad_0, pad_type = var_2587_pad_type_0, strides = var_2587_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("op_2587_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_69_cast_fp16 = mul(x = var_2581_cast_fp16, y = var_2587_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381152896)))];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_to_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2605_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2605_cast_fp16")];
            int32 var_2603 = const()[name = string("op_2603"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_57_cast_fp16 = concat(axis = var_2603, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2605_cast_fp16))[name = string("doubled_57_cast_fp16")];
            tensor<int32, [1]> out_29_axes_0 = const()[name = string("out_29_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_29_gamma_0_to_fp16 = const()[name = string("out_29_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406318784)))];
            fp16 var_2615_to_fp16 = const()[name = string("op_2615_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_2615_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2626_split_sizes_0 = const()[name = string("op_2626_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2626_axis_0 = const()[name = string("op_2626_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2626_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2626_cast_fp16_1 = split(axis = var_2626_axis_0, split_sizes = var_2626_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2626_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406327040)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2626_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2683_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2690_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2690_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2694_cast_fp16 = mul(x = x_71_cast_fp16, y = var_308_cast_fp16)[name = string("op_2694_cast_fp16")];
            tensor<int32, [2]> var_2695_split_sizes_0 = const()[name = string("op_2695_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2695_axis_0 = const()[name = string("op_2695_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2695_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2695_cast_fp16_1 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2695_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_2697_cast_fp16 = mul(x = var_2695_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2697_cast_fp16")];
            int32 var_2699 = const()[name = string("op_2699"), val = int32(-2)];
            bool var_2700_interleave_0 = const()[name = string("op_2700_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2700_cast_fp16 = concat(axis = var_2699, interleave = var_2700_interleave_0, values = (var_2697_cast_fp16, var_2695_cast_fp16_0))[name = string("op_2700_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2701_cast_fp16 = mul(x = var_2700_cast_fp16, y = var_315_cast_fp16)[name = string("op_2701_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_45_cast_fp16 = add(x = var_2694_cast_fp16, y = var_2701_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2707_cast_fp16 = mul(x = var_2683_cast_fp16, y = var_308_cast_fp16)[name = string("op_2707_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2708_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = var_2683_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_2710_cast_fp16 = mul(x = var_2708_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2710_cast_fp16")];
            int32 var_2712 = const()[name = string("op_2712"), val = int32(-2)];
            bool var_2713_interleave_0 = const()[name = string("op_2713_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2713_cast_fp16 = concat(axis = var_2712, interleave = var_2713_interleave_0, values = (var_2710_cast_fp16, var_2708_cast_fp16_0))[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2714_cast_fp16 = mul(x = var_2713_cast_fp16, y = var_315_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_75_cast_fp16 = add(x = var_2707_cast_fp16, y = var_2714_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_30)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_32_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_32 = read_state(input = key_cache)[name = string("coreml_update_state_32")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2690_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_31)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_33_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_33 = read_state(input = value_cache)[name = string("coreml_update_state_33")];
            tensor<int32, [4]> var_2784_begin_0 = const()[name = string("op_2784_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2784_end_0 = const()[name = string("op_2784_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2784_end_mask_0 = const()[name = string("op_2784_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2784_cast_fp16 = slice_by_index(begin = var_2784_begin_0, end = var_2784_end_0, end_mask = var_2784_end_mask_0, x = coreml_update_state_32)[name = string("op_2784_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2787_axis_0 = const()[name = string("op_2787_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_1 = split(axis = var_2787_axis_0, split_sizes = tile_14, x = var_2784_cast_fp16)[name = string("op_2787_cast_fp16")];
            tensor<int32, [4]> var_2794_begin_0 = const()[name = string("op_2794_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2794_end_0 = const()[name = string("op_2794_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2794_end_mask_0 = const()[name = string("op_2794_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2794_cast_fp16 = slice_by_index(begin = var_2794_begin_0, end = var_2794_end_0, end_mask = var_2794_end_mask_0, x = coreml_update_state_33)[name = string("op_2794_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2797_axis_0 = const()[name = string("op_2797_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_1 = split(axis = var_2797_axis_0, split_sizes = tile_15, x = var_2794_cast_fp16)[name = string("op_2797_cast_fp16")];
            tensor<int32, [2]> var_2800_split_sizes_0 = const()[name = string("op_2800_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2800_axis_0 = const()[name = string("op_2800_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2800_0, tensor<fp16, [1, 8, 128, 1]> var_2800_1 = split(axis = var_2800_axis_0, split_sizes = var_2800_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2800")];
            bool attn_weights_113_transpose_x_0 = const()[name = string("attn_weights_113_transpose_x_0"), val = bool(false)];
            bool attn_weights_113_transpose_y_0 = const()[name = string("attn_weights_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_113_cast_fp16 = matmul(transpose_x = attn_weights_113_transpose_x_0, transpose_y = attn_weights_113_transpose_y_0, x = var_2787_cast_fp16_0, y = var_2800_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2803_to_fp16 = const()[name = string("op_2803_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_115_cast_fp16 = mul(x = attn_weights_113_cast_fp16, y = var_2803_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            int32 var_2807 = const()[name = string("op_2807"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_119_cast_fp16 = softmax(axis = var_2807, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2813_transpose_x_1 = const()[name = string("op_2813_transpose_x_1"), val = bool(true)];
            bool var_2813_transpose_y_1 = const()[name = string("op_2813_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2813_cast_fp16 = matmul(transpose_x = var_2813_transpose_x_1, transpose_y = var_2813_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2797_cast_fp16_0)[name = string("op_2813_cast_fp16")];
            bool attn_weights_121_transpose_x_0 = const()[name = string("attn_weights_121_transpose_x_0"), val = bool(false)];
            bool attn_weights_121_transpose_y_0 = const()[name = string("attn_weights_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_0, transpose_y = attn_weights_121_transpose_y_0, x = var_2787_cast_fp16_1, y = var_2800_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2815_to_fp16 = const()[name = string("op_2815_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_2815_to_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_125_cast_fp16 = add(x = attn_weights_123_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_125_cast_fp16")];
            int32 var_2819 = const()[name = string("op_2819"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_127_cast_fp16 = softmax(axis = var_2819, x = attn_weights_125_cast_fp16)[name = string("attn_weights_127_cast_fp16")];
            bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(true)];
            bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_127_cast_fp16, y = var_2797_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2827 = const()[name = string("op_2827"), val = int32(1)];
            bool attn_output_59_interleave_0 = const()[name = string("attn_output_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_59_cast_fp16 = concat(axis = var_2827, interleave = attn_output_59_interleave_0, values = (var_2813_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2831_perm_0 = const()[name = string("op_2831_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_2831_cast_fp16 = transpose(perm = var_2831_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2831_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2864_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2864_cast_fp16")];
            int32 var_2862 = const()[name = string("op_2862"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_61_cast_fp16 = concat(axis = var_2862, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2864_cast_fp16))[name = string("doubled_61_cast_fp16")];
            tensor<int32, [1]> out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_31_gamma_0_to_fp16 = const()[name = string("out_31_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407375680)))];
            fp16 var_2874_to_fp16 = const()[name = string("op_2874_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_2874_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2885_split_sizes_0 = const()[name = string("op_2885_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2885_axis_0 = const()[name = string("op_2885_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2885_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2885_cast_fp16_1 = split(axis = var_2885_axis_0, split_sizes = var_2885_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2885_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2885_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2902_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2902_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407383936)))];
            tensor<int32, [2]> var_2908_strides_0 = const()[name = string("op_2908_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2908_pad_type_0 = const()[name = string("op_2908_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2908_pad_0 = const()[name = string("op_2908_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2908_dilations_0 = const()[name = string("op_2908_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2908_groups_0 = const()[name = string("op_2908_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2908_cast_fp16 = conv(dilations = var_2908_dilations_0, groups = var_2908_groups_0, pad = var_2908_pad_0, pad_type = var_2908_pad_type_0, strides = var_2908_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2885_cast_fp16_0)[name = string("op_2908_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_79_cast_fp16 = mul(x = var_2902_cast_fp16, y = var_2908_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432549824)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2926_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2926_cast_fp16")];
            int32 var_2924 = const()[name = string("op_2924"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_65_cast_fp16 = concat(axis = var_2924, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2926_cast_fp16))[name = string("doubled_65_cast_fp16")];
            tensor<int32, [1]> out_33_axes_0 = const()[name = string("out_33_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_33_gamma_0_to_fp16 = const()[name = string("out_33_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457715712)))];
            fp16 var_2936_to_fp16 = const()[name = string("op_2936_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_2936_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2947_split_sizes_0 = const()[name = string("op_2947_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2947_axis_0 = const()[name = string("op_2947_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2947_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2947_cast_fp16_1 = split(axis = var_2947_axis_0, split_sizes = var_2947_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2947_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457723968)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466112640)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467161280)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3004_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3004_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3011_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3011_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3015_cast_fp16 = mul(x = x_81_cast_fp16, y = var_308_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<int32, [2]> var_3016_split_sizes_0 = const()[name = string("op_3016_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3016_axis_0 = const()[name = string("op_3016_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3016_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3016_cast_fp16_1 = split(axis = var_3016_axis_0, split_sizes = var_3016_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3016_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_3018_cast_fp16 = mul(x = var_3016_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3018_cast_fp16")];
            int32 var_3020 = const()[name = string("op_3020"), val = int32(-2)];
            bool var_3021_interleave_0 = const()[name = string("op_3021_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3021_cast_fp16 = concat(axis = var_3020, interleave = var_3021_interleave_0, values = (var_3018_cast_fp16, var_3016_cast_fp16_0))[name = string("op_3021_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3022_cast_fp16 = mul(x = var_3021_cast_fp16, y = var_315_cast_fp16)[name = string("op_3022_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_51_cast_fp16 = add(x = var_3015_cast_fp16, y = var_3022_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3028_cast_fp16 = mul(x = var_3004_cast_fp16, y = var_308_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3029_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = var_3004_cast_fp16)[name = string("op_3029_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_3031_cast_fp16 = mul(x = var_3029_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3031_cast_fp16")];
            int32 var_3033 = const()[name = string("op_3033"), val = int32(-2)];
            bool var_3034_interleave_0 = const()[name = string("op_3034_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3034_cast_fp16 = concat(axis = var_3033, interleave = var_3034_interleave_0, values = (var_3031_cast_fp16, var_3029_cast_fp16_0))[name = string("op_3034_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3035_cast_fp16 = mul(x = var_3034_cast_fp16, y = var_315_cast_fp16)[name = string("op_3035_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_85_cast_fp16 = add(x = var_3028_cast_fp16, y = var_3035_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_32)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_34_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_34 = read_state(input = key_cache)[name = string("coreml_update_state_34")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3011_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_33)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_35_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_35 = read_state(input = value_cache)[name = string("coreml_update_state_35")];
            tensor<int32, [4]> var_3105_begin_0 = const()[name = string("op_3105_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3105_end_0 = const()[name = string("op_3105_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3105_end_mask_0 = const()[name = string("op_3105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3105_cast_fp16 = slice_by_index(begin = var_3105_begin_0, end = var_3105_end_0, end_mask = var_3105_end_mask_0, x = coreml_update_state_34)[name = string("op_3105_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3108_axis_0 = const()[name = string("op_3108_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_1 = split(axis = var_3108_axis_0, split_sizes = tile_16, x = var_3105_cast_fp16)[name = string("op_3108_cast_fp16")];
            tensor<int32, [4]> var_3115_begin_0 = const()[name = string("op_3115_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3115_end_0 = const()[name = string("op_3115_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3115_end_mask_0 = const()[name = string("op_3115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3115_cast_fp16 = slice_by_index(begin = var_3115_begin_0, end = var_3115_end_0, end_mask = var_3115_end_mask_0, x = coreml_update_state_35)[name = string("op_3115_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3118_axis_0 = const()[name = string("op_3118_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_1 = split(axis = var_3118_axis_0, split_sizes = tile_17, x = var_3115_cast_fp16)[name = string("op_3118_cast_fp16")];
            tensor<int32, [2]> var_3121_split_sizes_0 = const()[name = string("op_3121_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3121_axis_0 = const()[name = string("op_3121_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3121_0, tensor<fp16, [1, 8, 128, 1]> var_3121_1 = split(axis = var_3121_axis_0, split_sizes = var_3121_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3121")];
            bool attn_weights_129_transpose_x_0 = const()[name = string("attn_weights_129_transpose_x_0"), val = bool(false)];
            bool attn_weights_129_transpose_y_0 = const()[name = string("attn_weights_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_129_cast_fp16 = matmul(transpose_x = attn_weights_129_transpose_x_0, transpose_y = attn_weights_129_transpose_y_0, x = var_3108_cast_fp16_0, y = var_3121_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3124_to_fp16 = const()[name = string("op_3124_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_131_cast_fp16 = mul(x = attn_weights_129_cast_fp16, y = var_3124_to_fp16)[name = string("attn_weights_131_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_133_cast_fp16 = add(x = attn_weights_131_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_133_cast_fp16")];
            int32 var_3128 = const()[name = string("op_3128"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_135_cast_fp16 = softmax(axis = var_3128, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3134_transpose_x_1 = const()[name = string("op_3134_transpose_x_1"), val = bool(true)];
            bool var_3134_transpose_y_1 = const()[name = string("op_3134_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3134_cast_fp16 = matmul(transpose_x = var_3134_transpose_x_1, transpose_y = var_3134_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3118_cast_fp16_0)[name = string("op_3134_cast_fp16")];
            bool attn_weights_137_transpose_x_0 = const()[name = string("attn_weights_137_transpose_x_0"), val = bool(false)];
            bool attn_weights_137_transpose_y_0 = const()[name = string("attn_weights_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_137_cast_fp16 = matmul(transpose_x = attn_weights_137_transpose_x_0, transpose_y = attn_weights_137_transpose_y_0, x = var_3108_cast_fp16_1, y = var_3121_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3136_to_fp16 = const()[name = string("op_3136_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_139_cast_fp16 = mul(x = attn_weights_137_cast_fp16, y = var_3136_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            int32 var_3140 = const()[name = string("op_3140"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_cast_fp16 = softmax(axis = var_3140, x = attn_weights_141_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(true)];
            bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_cast_fp16, y = var_3118_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3148 = const()[name = string("op_3148"), val = int32(1)];
            bool attn_output_67_interleave_0 = const()[name = string("attn_output_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_67_cast_fp16 = concat(axis = var_3148, interleave = attn_output_67_interleave_0, values = (var_3134_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3152_perm_0 = const()[name = string("op_3152_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_3152_cast_fp16 = transpose(perm = var_3152_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_cast_fp16 = reshape(shape = concat_107x, x = var_3152_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468209920)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3185_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3185_cast_fp16")];
            int32 var_3183 = const()[name = string("op_3183"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_69_cast_fp16 = concat(axis = var_3183, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3185_cast_fp16))[name = string("doubled_69_cast_fp16")];
            tensor<int32, [1]> out_35_axes_0 = const()[name = string("out_35_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_35_gamma_0_to_fp16 = const()[name = string("out_35_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476598592)))];
            fp16 var_3195_to_fp16 = const()[name = string("op_3195_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_3195_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3206_split_sizes_0 = const()[name = string("op_3206_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3206_axis_0 = const()[name = string("op_3206_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3206_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3206_cast_fp16_1 = split(axis = var_3206_axis_0, split_sizes = var_3206_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3206_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476606848)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_8_mlp_gate_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3223_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3223_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501772736)))];
            tensor<int32, [2]> var_3229_strides_0 = const()[name = string("op_3229_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3229_pad_type_0 = const()[name = string("op_3229_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3229_pad_0 = const()[name = string("op_3229_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3229_dilations_0 = const()[name = string("op_3229_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3229_groups_0 = const()[name = string("op_3229_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3229_cast_fp16 = conv(dilations = var_3229_dilations_0, groups = var_3229_groups_0, pad = var_3229_pad_0, pad_type = var_3229_pad_type_0, strides = var_3229_strides_0, weight = layers_8_mlp_up_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("op_3229_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_cast_fp16 = mul(x = var_3223_cast_fp16, y = var_3229_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526938624)))];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3247_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3247_cast_fp16")];
            int32 var_3245 = const()[name = string("op_3245"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_73_cast_fp16 = concat(axis = var_3245, interleave = doubled_73_interleave_0, values = (hidden_states_cast_fp16, var_3247_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_axes_0 = const()[name = string("out_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_gamma_0_to_fp16 = const()[name = string("out_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552104512)))];
            fp16 var_3257_to_fp16 = const()[name = string("op_3257_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3257_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3268_split_sizes_0 = const()[name = string("op_3268_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3268_axis_0 = const()[name = string("op_3268_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states, tensor<fp16, [1, 2048, 1, 1]> var_3268_cast_fp16_1 = split(axis = var_3268_axis_0, split_sizes = var_3268_split_sizes_0, x = out_cast_fp16)[name = string("op_3268_cast_fp16")];
        } -> (hidden_states);
    func length_128<ios18>(tensor<fp32, [1, 2048, 1, 128]> inputs_embeds, state<tensor<fp16, [9, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [128]> position_index_seed, state<tensor<fp16, [9, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725952))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17321280))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17308928))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17327488))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29922816))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29910464))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29929024))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42516160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42512000))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42518272))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243840))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243264))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47244160))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51442688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51438528))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444800))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64040128))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64027776))))[name = string("layers_1_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64046336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76633472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76629312))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76635584))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80834112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80829952))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80836224))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361152))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81360576))))[name = string("layers_2_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361472))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85560000))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85555840))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85562112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98145088))))[name = string("layers_2_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110758976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110746624))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110765184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123352320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123348160))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123354432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127552960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127548800))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127555072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128079424))))[name = string("layers_3_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080320))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140675648))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140663296))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140681856))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153277184))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153264832))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153283392))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165870528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165866368))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165872640))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170071168))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170067008))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170073280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170597632))))[name = string("layers_4_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174797056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174792896))))[name = string("layers_4_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174799168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187394496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187382144))))[name = string("layers_4_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187400704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199996032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199983680))))[name = string("layers_4_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_4_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200002240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212589376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212585216))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212591488))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216790016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216785856))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216792128))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317056))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217316480))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317376))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221515904))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221511744))))[name = string("layers_5_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221518016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234113344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234100992))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234119552))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246714880))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246702528))))[name = string("layers_5_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_5_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246721088))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259308224))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259304064))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259310336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263508864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263504704))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263510976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035328))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264036224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268234752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268230592))))[name = string("layers_6_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268236864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280832192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280819840))))[name = string("layers_6_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280838400))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293433728))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293421376))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293439936))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297638464))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297634304))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297640576))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165504))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298164928))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165824))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302364352))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302360192))))[name = string("layers_7_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302366464))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314961792))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314949440))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(128)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [128]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_280 = const()[name = string("op_280"), val = int32(0)];
            bool var_282_exclusive_0 = const()[name = string("op_282_exclusive_0"), val = bool(false)];
            bool var_282_reverse_0 = const()[name = string("op_282_reverse_0"), val = bool(false)];
            tensor<fp16, [128]> var_282_cast_fp16 = cumsum(axis = var_280, exclusive = var_282_exclusive_0, reverse = var_282_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_282_cast_fp16")];
            fp16 var_284_promoted_to_fp16 = const()[name = string("op_284_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [128]> position_offsets_cast_fp16 = sub(x = var_282_cast_fp16, y = var_284_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 128]> var_287_cast_fp16 = expand_dims(axes = var_287_axes_0, x = position_offsets_cast_fp16)[name = string("op_287_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_27")];
            tensor<fp16, [1, 128]> position_ids_1_cast_fp16 = add(x = var_287_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 128]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_26")];
            tensor<bool, [1, 128]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 128]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 128]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314968000)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_12_y_0 = const()[name = string("greater_equal_12_y_0"), val = int32(0)];
            tensor<bool, [1, 128]> greater_equal_12 = greater_equal(x = select_0, y = greater_equal_12_y_0)[name = string("greater_equal_12")];
            int32 slice_by_index_12 = const()[name = string("slice_by_index_12"), val = int32(32768)];
            tensor<int32, [1, 128]> add_12 = add(x = select_0, y = slice_by_index_12)[name = string("add_12")];
            tensor<int32, [1, 128]> select_12 = select(a = select_0, b = add_12, cond = greater_equal_12)[name = string("select_12")];
            int32 cos_1_cast_fp16_axis_6 = const()[name = string("cos_1_cast_fp16_axis_6"), val = int32(0)];
            tensor<fp16, [1, 128, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_6, batch_dims = cos_1_batch_dims_0, indices = select_12, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323356672)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_6 = const()[name = string("sin_1_cast_fp16_axis_6"), val = int32(0)];
            tensor<fp16, [1, 128, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_6, batch_dims = sin_1_batch_dims_0, indices = select_12, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_306_perm_0 = const()[name = string("op_306_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_308_axes_0 = const()[name = string("op_308_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 128]> var_306_cast_fp16 = transpose(perm = var_306_perm_0, x = cos_1_cast_fp16)[name = string("transpose_209")];
            tensor<fp16, [1, 1, 128, 128]> var_308_cast_fp16 = expand_dims(axes = var_308_axes_0, x = var_306_cast_fp16)[name = string("op_308_cast_fp16")];
            tensor<int32, [3]> var_313_perm_0 = const()[name = string("op_313_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 128]> var_313_cast_fp16 = transpose(perm = var_313_perm_0, x = sin_1_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [1, 1, 128, 128]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = var_313_cast_fp16)[name = string("op_315_cast_fp16")];
            tensor<int32, [1]> var_334_axes_0 = const()[name = string("op_334_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 128, 1]> var_334 = expand_dims(axes = var_334_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_334")];
            tensor<int32, [1, 1, 2048]> var_327 = const()[name = string("op_327"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331745344)))];
            tensor<bool, [1, 128, 2048]> var_335 = greater(x = var_327, y = var_334)[name = string("op_335")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_342_axes_0 = const()[name = string("op_342_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 2048]> var_335_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_335)[name = string("cast_25")];
            tensor<fp16, [1, 1, 128, 2048]> var_342_cast_fp16 = expand_dims(axes = var_342_axes_0, x = var_335_to_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_346_promoted_to_fp16 = const()[name = string("op_346_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 128]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_342_cast_fp16)[name = string("transpose_207")];
            tensor<bool, [1, 1, 2048, 128]> var_347_cast_fp16 = equal(x = mask_cast_fp16, y = var_346_promoted_to_fp16)[name = string("op_347_cast_fp16")];
            fp16 var_348_to_fp16 = const()[name = string("op_348_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 128]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_348_to_fp16, cond = var_347_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_24")];
            tensor<fp16, [1, 2048, 1, 128]> var_358_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_358_cast_fp16")];
            int32 var_356 = const()[name = string("op_356"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_1_cast_fp16 = concat(axis = var_356, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_358_cast_fp16))[name = string("doubled_1_cast_fp16")];
            tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_1_gamma_0_to_fp16 = const()[name = string("out_1_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331753600)))];
            fp16 var_368_to_fp16 = const()[name = string("op_368_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_368_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_379_split_sizes_0 = const()[name = string("op_379_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_379_axis_0 = const()[name = string("op_379_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_379_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_379_cast_fp16_1 = split(axis = var_379_axis_0, split_sizes = var_379_split_sizes_0, x = out_1_cast_fp16)[name = string("op_379_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331761856)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_379_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_436_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_436_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_443_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_443_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_447_cast_fp16 = mul(x = x_1_cast_fp16, y = var_308_cast_fp16)[name = string("op_447_cast_fp16")];
            tensor<int32, [2]> var_448_split_sizes_0 = const()[name = string("op_448_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_448_axis_0 = const()[name = string("op_448_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_448_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_448_cast_fp16_1 = split(axis = var_448_axis_0, split_sizes = var_448_split_sizes_0, x = x_1_cast_fp16)[name = string("op_448_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_450_cast_fp16 = mul(x = var_448_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_450_cast_fp16")];
            int32 var_452 = const()[name = string("op_452"), val = int32(-2)];
            bool var_453_interleave_0 = const()[name = string("op_453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_453_cast_fp16 = concat(axis = var_452, interleave = var_453_interleave_0, values = (var_450_cast_fp16, var_448_cast_fp16_0))[name = string("op_453_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_454_cast_fp16 = mul(x = var_453_cast_fp16, y = var_315_cast_fp16)[name = string("op_454_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_3_cast_fp16 = add(x = var_447_cast_fp16, y = var_454_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_460_cast_fp16 = mul(x = var_436_cast_fp16, y = var_308_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [2]> var_461_split_sizes_0 = const()[name = string("op_461_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_461_axis_0 = const()[name = string("op_461_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_461_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_461_cast_fp16_1 = split(axis = var_461_axis_0, split_sizes = var_461_split_sizes_0, x = var_436_cast_fp16)[name = string("op_461_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_463_cast_fp16 = mul(x = var_461_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_463_cast_fp16")];
            int32 var_465 = const()[name = string("op_465"), val = int32(-2)];
            bool var_466_interleave_0 = const()[name = string("op_466_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_466_cast_fp16 = concat(axis = var_465, interleave = var_466_interleave_0, values = (var_463_cast_fp16, var_461_cast_fp16_0))[name = string("op_466_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_467_cast_fp16 = mul(x = var_466_cast_fp16, y = var_315_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_5_cast_fp16 = add(x = var_460_cast_fp16, y = var_467_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_108_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_108 = read_state(input = key_cache)[name = string("coreml_update_state_108")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_443_cast_fp16)[name = string("transpose_205")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_109_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_109 = read_state(input = value_cache)[name = string("coreml_update_state_109")];
            tensor<int32, [4]> var_537_begin_0 = const()[name = string("op_537_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_537_end_0 = const()[name = string("op_537_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_537_end_mask_0 = const()[name = string("op_537_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_537_cast_fp16 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = coreml_update_state_108)[name = string("op_537_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_540_axis_0 = const()[name = string("op_540_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_1 = split(axis = var_540_axis_0, split_sizes = tile_0, x = var_537_cast_fp16)[name = string("op_540_cast_fp16")];
            tensor<int32, [4]> var_547_begin_0 = const()[name = string("op_547_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_547_end_0 = const()[name = string("op_547_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_547_end_mask_0 = const()[name = string("op_547_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_547_cast_fp16 = slice_by_index(begin = var_547_begin_0, end = var_547_end_0, end_mask = var_547_end_mask_0, x = coreml_update_state_109)[name = string("op_547_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_550_axis_0 = const()[name = string("op_550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_1 = split(axis = var_550_axis_0, split_sizes = tile_1, x = var_547_cast_fp16)[name = string("op_550_cast_fp16")];
            tensor<int32, [2]> var_553_split_sizes_0 = const()[name = string("op_553_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_553_axis_0 = const()[name = string("op_553_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_553_0, tensor<fp16, [1, 8, 128, 128]> var_553_1 = split(axis = var_553_axis_0, split_sizes = var_553_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_553")];
            bool attn_weights_1_transpose_x_0 = const()[name = string("attn_weights_1_transpose_x_0"), val = bool(false)];
            bool attn_weights_1_transpose_y_0 = const()[name = string("attn_weights_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_540_cast_fp16_0, y = var_553_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_556_to_fp16 = const()[name = string("op_556_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_556_to_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_5_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
            int32 var_560 = const()[name = string("op_560"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_7_cast_fp16 = softmax(axis = var_560, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_566_transpose_x_1 = const()[name = string("op_566_transpose_x_1"), val = bool(true)];
            bool var_566_transpose_y_1 = const()[name = string("op_566_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_566_cast_fp16 = matmul(transpose_x = var_566_transpose_x_1, transpose_y = var_566_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_550_cast_fp16_0)[name = string("op_566_cast_fp16")];
            bool attn_weights_9_transpose_x_0 = const()[name = string("attn_weights_9_transpose_x_0"), val = bool(false)];
            bool attn_weights_9_transpose_y_0 = const()[name = string("attn_weights_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_9_cast_fp16 = matmul(transpose_x = attn_weights_9_transpose_x_0, transpose_y = attn_weights_9_transpose_y_0, x = var_540_cast_fp16_1, y = var_553_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_568_to_fp16 = const()[name = string("op_568_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_11_cast_fp16 = mul(x = attn_weights_9_cast_fp16, y = var_568_to_fp16)[name = string("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_13_cast_fp16 = add(x = attn_weights_11_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_13_cast_fp16")];
            int32 var_572 = const()[name = string("op_572"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_15_cast_fp16 = softmax(axis = var_572, x = attn_weights_13_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(true)];
            bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_15_cast_fp16, y = var_550_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_580 = const()[name = string("op_580"), val = int32(1)];
            bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_3_cast_fp16 = concat(axis = var_580, interleave = attn_output_3_interleave_0, values = (var_566_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_584_perm_0 = const()[name = string("op_584_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_584_cast_fp16 = transpose(perm = var_584_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_204")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_584_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332810496)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_617_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_617_cast_fp16")];
            int32 var_615 = const()[name = string("op_615"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_5_cast_fp16 = concat(axis = var_615, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_617_cast_fp16))[name = string("doubled_5_cast_fp16")];
            tensor<int32, [1]> out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_3_gamma_0_to_fp16 = const()[name = string("out_3_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341199168)))];
            fp16 var_627_to_fp16 = const()[name = string("op_627_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_627_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_638_split_sizes_0 = const()[name = string("op_638_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_638_axis_0 = const()[name = string("op_638_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_638_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_638_cast_fp16_1 = split(axis = var_638_axis_0, split_sizes = var_638_split_sizes_0, x = out_3_cast_fp16)[name = string("op_638_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_655_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_655_cast_fp16")];
            tensor<int32, [2]> var_661_strides_0 = const()[name = string("op_661_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_661_pad_type_0 = const()[name = string("op_661_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_661_pad_0 = const()[name = string("op_661_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_661_dilations_0 = const()[name = string("op_661_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_661_groups_0 = const()[name = string("op_661_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_661_cast_fp16 = conv(dilations = var_661_dilations_0, groups = var_661_groups_0, pad = var_661_pad_0, pad_type = var_661_pad_type_0, strides = var_661_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("op_661_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_9_cast_fp16 = mul(x = var_655_cast_fp16, y = var_661_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_679_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_679_cast_fp16")];
            int32 var_677 = const()[name = string("op_677"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_9_cast_fp16 = concat(axis = var_677, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_679_cast_fp16))[name = string("doubled_9_cast_fp16")];
            tensor<int32, [1]> out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_5_gamma_0_to_fp16 = const()[name = string("out_5_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341207424)))];
            fp16 var_689_to_fp16 = const()[name = string("op_689_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_689_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_700_split_sizes_0 = const()[name = string("op_700_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_700_axis_0 = const()[name = string("op_700_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_700_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_700_cast_fp16_1 = split(axis = var_700_axis_0, split_sizes = var_700_split_sizes_0, x = out_5_cast_fp16)[name = string("op_700_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341215680)))];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = var_700_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_757_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_757_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_764_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_764_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_768_cast_fp16 = mul(x = x_11_cast_fp16, y = var_308_cast_fp16)[name = string("op_768_cast_fp16")];
            tensor<int32, [2]> var_769_split_sizes_0 = const()[name = string("op_769_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_769_axis_0 = const()[name = string("op_769_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_769_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_769_cast_fp16_1 = split(axis = var_769_axis_0, split_sizes = var_769_split_sizes_0, x = x_11_cast_fp16)[name = string("op_769_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_771_cast_fp16 = mul(x = var_769_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_771_cast_fp16")];
            int32 var_773 = const()[name = string("op_773"), val = int32(-2)];
            bool var_774_interleave_0 = const()[name = string("op_774_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_774_cast_fp16 = concat(axis = var_773, interleave = var_774_interleave_0, values = (var_771_cast_fp16, var_769_cast_fp16_0))[name = string("op_774_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_775_cast_fp16 = mul(x = var_774_cast_fp16, y = var_315_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_9_cast_fp16 = add(x = var_768_cast_fp16, y = var_775_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_781_cast_fp16 = mul(x = var_757_cast_fp16, y = var_308_cast_fp16)[name = string("op_781_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_782_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = var_757_cast_fp16)[name = string("op_782_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_784_cast_fp16 = mul(x = var_782_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_784_cast_fp16")];
            int32 var_786 = const()[name = string("op_786"), val = int32(-2)];
            bool var_787_interleave_0 = const()[name = string("op_787_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_787_cast_fp16 = concat(axis = var_786, interleave = var_787_interleave_0, values = (var_784_cast_fp16, var_782_cast_fp16_0))[name = string("op_787_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_788_cast_fp16 = mul(x = var_787_cast_fp16, y = var_315_cast_fp16)[name = string("op_788_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_15_cast_fp16 = add(x = var_781_cast_fp16, y = var_788_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_203")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_108)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_110_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_110 = read_state(input = key_cache)[name = string("coreml_update_state_110")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_764_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_109)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_111_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_111 = read_state(input = value_cache)[name = string("coreml_update_state_111")];
            tensor<int32, [4]> var_858_begin_0 = const()[name = string("op_858_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_858_end_0 = const()[name = string("op_858_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_858_end_mask_0 = const()[name = string("op_858_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_858_cast_fp16 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = coreml_update_state_110)[name = string("op_858_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_861_axis_0 = const()[name = string("op_861_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_1 = split(axis = var_861_axis_0, split_sizes = tile_2, x = var_858_cast_fp16)[name = string("op_861_cast_fp16")];
            tensor<int32, [4]> var_868_begin_0 = const()[name = string("op_868_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_868_end_0 = const()[name = string("op_868_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_868_end_mask_0 = const()[name = string("op_868_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_868_cast_fp16 = slice_by_index(begin = var_868_begin_0, end = var_868_end_0, end_mask = var_868_end_mask_0, x = coreml_update_state_111)[name = string("op_868_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_871_axis_0 = const()[name = string("op_871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_1 = split(axis = var_871_axis_0, split_sizes = tile_3, x = var_868_cast_fp16)[name = string("op_871_cast_fp16")];
            tensor<int32, [2]> var_874_split_sizes_0 = const()[name = string("op_874_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_874_axis_0 = const()[name = string("op_874_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_874_0, tensor<fp16, [1, 8, 128, 128]> var_874_1 = split(axis = var_874_axis_0, split_sizes = var_874_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_874")];
            bool attn_weights_17_transpose_x_0 = const()[name = string("attn_weights_17_transpose_x_0"), val = bool(false)];
            bool attn_weights_17_transpose_y_0 = const()[name = string("attn_weights_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_17_cast_fp16 = matmul(transpose_x = attn_weights_17_transpose_x_0, transpose_y = attn_weights_17_transpose_y_0, x = var_861_cast_fp16_0, y = var_874_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_877_to_fp16 = const()[name = string("op_877_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_19_cast_fp16 = mul(x = attn_weights_17_cast_fp16, y = var_877_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            int32 var_881 = const()[name = string("op_881"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_23_cast_fp16 = softmax(axis = var_881, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_887_transpose_x_1 = const()[name = string("op_887_transpose_x_1"), val = bool(true)];
            bool var_887_transpose_y_1 = const()[name = string("op_887_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_887_cast_fp16 = matmul(transpose_x = var_887_transpose_x_1, transpose_y = var_887_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_871_cast_fp16_0)[name = string("op_887_cast_fp16")];
            bool attn_weights_25_transpose_x_0 = const()[name = string("attn_weights_25_transpose_x_0"), val = bool(false)];
            bool attn_weights_25_transpose_y_0 = const()[name = string("attn_weights_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_0, transpose_y = attn_weights_25_transpose_y_0, x = var_861_cast_fp16_1, y = var_874_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_889_to_fp16 = const()[name = string("op_889_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_889_to_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_29_cast_fp16 = add(x = attn_weights_27_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
            int32 var_893 = const()[name = string("op_893"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_31_cast_fp16 = softmax(axis = var_893, x = attn_weights_29_cast_fp16)[name = string("attn_weights_31_cast_fp16")];
            bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(true)];
            bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_31_cast_fp16, y = var_871_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_901 = const()[name = string("op_901"), val = int32(1)];
            bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_11_cast_fp16 = concat(axis = var_901, interleave = attn_output_11_interleave_0, values = (var_887_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_905_perm_0 = const()[name = string("op_905_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_905_cast_fp16 = transpose(perm = var_905_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_201")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_905_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_938_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_938_cast_fp16")];
            int32 var_936 = const()[name = string("op_936"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_13_cast_fp16 = concat(axis = var_936, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_938_cast_fp16))[name = string("doubled_13_cast_fp16")];
            tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_7_gamma_0_to_fp16 = const()[name = string("out_7_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342264320)))];
            fp16 var_948_to_fp16 = const()[name = string("op_948_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_948_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_959_split_sizes_0 = const()[name = string("op_959_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_959_axis_0 = const()[name = string("op_959_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_959_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_959_cast_fp16_1 = split(axis = var_959_axis_0, split_sizes = var_959_split_sizes_0, x = out_7_cast_fp16)[name = string("op_959_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_959_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_976_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_976_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342272576)))];
            tensor<int32, [2]> var_982_strides_0 = const()[name = string("op_982_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_982_pad_type_0 = const()[name = string("op_982_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_982_pad_0 = const()[name = string("op_982_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_982_dilations_0 = const()[name = string("op_982_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_982_groups_0 = const()[name = string("op_982_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_982_cast_fp16 = conv(dilations = var_982_dilations_0, groups = var_982_groups_0, pad = var_982_pad_0, pad_type = var_982_pad_type_0, strides = var_982_strides_0, weight = layers_1_mlp_up_proj_weight_to_fp16, x = var_959_cast_fp16_0)[name = string("op_982_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_19_cast_fp16 = mul(x = var_976_cast_fp16, y = var_982_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1000_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1000_cast_fp16")];
            int32 var_998 = const()[name = string("op_998"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_17_cast_fp16 = concat(axis = var_998, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1000_cast_fp16))[name = string("doubled_17_cast_fp16")];
            tensor<int32, [1]> out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_9_gamma_0_to_fp16 = const()[name = string("out_9_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367438464)))];
            fp16 var_1010_to_fp16 = const()[name = string("op_1010_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1010_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1021_split_sizes_0 = const()[name = string("op_1021_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1021_axis_0 = const()[name = string("op_1021_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1021_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1021_cast_fp16_1 = split(axis = var_1021_axis_0, split_sizes = var_1021_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1021_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367446720)))];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = var_1021_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_1078_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1078_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_1085_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1085_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1089_cast_fp16 = mul(x = x_21_cast_fp16, y = var_308_cast_fp16)[name = string("op_1089_cast_fp16")];
            tensor<int32, [2]> var_1090_split_sizes_0 = const()[name = string("op_1090_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1090_axis_0 = const()[name = string("op_1090_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_1090_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_1090_cast_fp16_1 = split(axis = var_1090_axis_0, split_sizes = var_1090_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1090_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_1092_cast_fp16 = mul(x = var_1090_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            int32 var_1094 = const()[name = string("op_1094"), val = int32(-2)];
            bool var_1095_interleave_0 = const()[name = string("op_1095_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_1095_cast_fp16 = concat(axis = var_1094, interleave = var_1095_interleave_0, values = (var_1092_cast_fp16, var_1090_cast_fp16_0))[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1096_cast_fp16 = mul(x = var_1095_cast_fp16, y = var_315_cast_fp16)[name = string("op_1096_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_15_cast_fp16 = add(x = var_1089_cast_fp16, y = var_1096_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1102_cast_fp16 = mul(x = var_1078_cast_fp16, y = var_308_cast_fp16)[name = string("op_1102_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_1103_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = var_1078_cast_fp16)[name = string("op_1103_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_1105_cast_fp16 = mul(x = var_1103_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1105_cast_fp16")];
            int32 var_1107 = const()[name = string("op_1107"), val = int32(-2)];
            bool var_1108_interleave_0 = const()[name = string("op_1108_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_1108_cast_fp16 = concat(axis = var_1107, interleave = var_1108_interleave_0, values = (var_1105_cast_fp16, var_1103_cast_fp16_0))[name = string("op_1108_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1109_cast_fp16 = mul(x = var_1108_cast_fp16, y = var_315_cast_fp16)[name = string("op_1109_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_25_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1109_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_200")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_110)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_112_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_112 = read_state(input = key_cache)[name = string("coreml_update_state_112")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1085_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_111)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_113_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_113 = read_state(input = value_cache)[name = string("coreml_update_state_113")];
            tensor<int32, [4]> var_1179_begin_0 = const()[name = string("op_1179_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1179_end_0 = const()[name = string("op_1179_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1179_end_mask_0 = const()[name = string("op_1179_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1179_cast_fp16 = slice_by_index(begin = var_1179_begin_0, end = var_1179_end_0, end_mask = var_1179_end_mask_0, x = coreml_update_state_112)[name = string("op_1179_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1182_axis_0 = const()[name = string("op_1182_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_1 = split(axis = var_1182_axis_0, split_sizes = tile_4, x = var_1179_cast_fp16)[name = string("op_1182_cast_fp16")];
            tensor<int32, [4]> var_1189_begin_0 = const()[name = string("op_1189_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1189_end_0 = const()[name = string("op_1189_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1189_end_mask_0 = const()[name = string("op_1189_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1189_cast_fp16 = slice_by_index(begin = var_1189_begin_0, end = var_1189_end_0, end_mask = var_1189_end_mask_0, x = coreml_update_state_113)[name = string("op_1189_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1192_axis_0 = const()[name = string("op_1192_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_1 = split(axis = var_1192_axis_0, split_sizes = tile_5, x = var_1189_cast_fp16)[name = string("op_1192_cast_fp16")];
            tensor<int32, [2]> var_1195_split_sizes_0 = const()[name = string("op_1195_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1195_axis_0 = const()[name = string("op_1195_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1195_0, tensor<fp16, [1, 8, 128, 128]> var_1195_1 = split(axis = var_1195_axis_0, split_sizes = var_1195_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1195")];
            bool attn_weights_33_transpose_x_0 = const()[name = string("attn_weights_33_transpose_x_0"), val = bool(false)];
            bool attn_weights_33_transpose_y_0 = const()[name = string("attn_weights_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_33_cast_fp16 = matmul(transpose_x = attn_weights_33_transpose_x_0, transpose_y = attn_weights_33_transpose_y_0, x = var_1182_cast_fp16_0, y = var_1195_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1198_to_fp16 = const()[name = string("op_1198_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_35_cast_fp16 = mul(x = attn_weights_33_cast_fp16, y = var_1198_to_fp16)[name = string("attn_weights_35_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_37_cast_fp16 = add(x = attn_weights_35_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_37_cast_fp16")];
            int32 var_1202 = const()[name = string("op_1202"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_39_cast_fp16 = softmax(axis = var_1202, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1208_transpose_x_1 = const()[name = string("op_1208_transpose_x_1"), val = bool(true)];
            bool var_1208_transpose_y_1 = const()[name = string("op_1208_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1208_cast_fp16 = matmul(transpose_x = var_1208_transpose_x_1, transpose_y = var_1208_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1192_cast_fp16_0)[name = string("op_1208_cast_fp16")];
            bool attn_weights_41_transpose_x_0 = const()[name = string("attn_weights_41_transpose_x_0"), val = bool(false)];
            bool attn_weights_41_transpose_y_0 = const()[name = string("attn_weights_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_41_cast_fp16 = matmul(transpose_x = attn_weights_41_transpose_x_0, transpose_y = attn_weights_41_transpose_y_0, x = var_1182_cast_fp16_1, y = var_1195_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1210_to_fp16 = const()[name = string("op_1210_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_43_cast_fp16 = mul(x = attn_weights_41_cast_fp16, y = var_1210_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            int32 var_1214 = const()[name = string("op_1214"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_47_cast_fp16 = softmax(axis = var_1214, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")];
            bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(true)];
            bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_47_cast_fp16, y = var_1192_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1222 = const()[name = string("op_1222"), val = int32(1)];
            bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_19_cast_fp16 = concat(axis = var_1222, interleave = attn_output_19_interleave_0, values = (var_1208_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1226_perm_0 = const()[name = string("op_1226_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_1226_cast_fp16 = transpose(perm = var_1226_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_198")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1226_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1259_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            int32 var_1257 = const()[name = string("op_1257"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_21_cast_fp16 = concat(axis = var_1257, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1259_cast_fp16))[name = string("doubled_21_cast_fp16")];
            tensor<int32, [1]> out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_11_gamma_0_to_fp16 = const()[name = string("out_11_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368495360)))];
            fp16 var_1269_to_fp16 = const()[name = string("op_1269_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1269_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1280_split_sizes_0 = const()[name = string("op_1280_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1280_axis_0 = const()[name = string("op_1280_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1280_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1280_cast_fp16_1 = split(axis = var_1280_axis_0, split_sizes = var_1280_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1280_cast_fp16")];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_1297_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1297_cast_fp16")];
            tensor<int32, [2]> var_1303_strides_0 = const()[name = string("op_1303_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1303_pad_type_0 = const()[name = string("op_1303_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1303_pad_0 = const()[name = string("op_1303_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1303_dilations_0 = const()[name = string("op_1303_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1303_groups_0 = const()[name = string("op_1303_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_1303_cast_fp16 = conv(dilations = var_1303_dilations_0, groups = var_1303_groups_0, pad = var_1303_pad_0, pad_type = var_1303_pad_type_0, strides = var_1303_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("op_1303_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_29_cast_fp16 = mul(x = var_1297_cast_fp16, y = var_1303_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1321_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1321_cast_fp16")];
            int32 var_1319 = const()[name = string("op_1319"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_25_cast_fp16 = concat(axis = var_1319, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1321_cast_fp16))[name = string("doubled_25_cast_fp16")];
            tensor<int32, [1]> out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_13_gamma_0_to_fp16 = const()[name = string("out_13_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368503616)))];
            fp16 var_1331_to_fp16 = const()[name = string("op_1331_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1331_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1342_split_sizes_0 = const()[name = string("op_1342_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1342_axis_0 = const()[name = string("op_1342_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1342_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1342_cast_fp16_1 = split(axis = var_1342_axis_0, split_sizes = var_1342_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1342_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368511872)))];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = var_1342_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_1399_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_1406_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1406_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1410_cast_fp16 = mul(x = x_31_cast_fp16, y = var_308_cast_fp16)[name = string("op_1410_cast_fp16")];
            tensor<int32, [2]> var_1411_split_sizes_0 = const()[name = string("op_1411_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1411_axis_0 = const()[name = string("op_1411_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_1411_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_1411_cast_fp16_1 = split(axis = var_1411_axis_0, split_sizes = var_1411_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1411_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_1413_cast_fp16 = mul(x = var_1411_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1413_cast_fp16")];
            int32 var_1415 = const()[name = string("op_1415"), val = int32(-2)];
            bool var_1416_interleave_0 = const()[name = string("op_1416_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_1416_cast_fp16 = concat(axis = var_1415, interleave = var_1416_interleave_0, values = (var_1413_cast_fp16, var_1411_cast_fp16_0))[name = string("op_1416_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1417_cast_fp16 = mul(x = var_1416_cast_fp16, y = var_315_cast_fp16)[name = string("op_1417_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_21_cast_fp16 = add(x = var_1410_cast_fp16, y = var_1417_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1423_cast_fp16 = mul(x = var_1399_cast_fp16, y = var_308_cast_fp16)[name = string("op_1423_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_1424_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = var_1399_cast_fp16)[name = string("op_1424_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_1426_cast_fp16 = mul(x = var_1424_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1426_cast_fp16")];
            int32 var_1428 = const()[name = string("op_1428"), val = int32(-2)];
            bool var_1429_interleave_0 = const()[name = string("op_1429_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_1429_cast_fp16 = concat(axis = var_1428, interleave = var_1429_interleave_0, values = (var_1426_cast_fp16, var_1424_cast_fp16_0))[name = string("op_1429_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1430_cast_fp16 = mul(x = var_1429_cast_fp16, y = var_315_cast_fp16)[name = string("op_1430_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_35_cast_fp16 = add(x = var_1423_cast_fp16, y = var_1430_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_112)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_114_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_114 = read_state(input = key_cache)[name = string("coreml_update_state_114")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1406_cast_fp16)[name = string("transpose_196")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_113)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_115_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_115 = read_state(input = value_cache)[name = string("coreml_update_state_115")];
            tensor<int32, [4]> var_1500_begin_0 = const()[name = string("op_1500_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1500_end_0 = const()[name = string("op_1500_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1500_end_mask_0 = const()[name = string("op_1500_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1500_cast_fp16 = slice_by_index(begin = var_1500_begin_0, end = var_1500_end_0, end_mask = var_1500_end_mask_0, x = coreml_update_state_114)[name = string("op_1500_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1503_axis_0 = const()[name = string("op_1503_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_1 = split(axis = var_1503_axis_0, split_sizes = tile_6, x = var_1500_cast_fp16)[name = string("op_1503_cast_fp16")];
            tensor<int32, [4]> var_1510_begin_0 = const()[name = string("op_1510_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1510_end_0 = const()[name = string("op_1510_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1510_end_mask_0 = const()[name = string("op_1510_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1510_cast_fp16 = slice_by_index(begin = var_1510_begin_0, end = var_1510_end_0, end_mask = var_1510_end_mask_0, x = coreml_update_state_115)[name = string("op_1510_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1513_axis_0 = const()[name = string("op_1513_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_1 = split(axis = var_1513_axis_0, split_sizes = tile_7, x = var_1510_cast_fp16)[name = string("op_1513_cast_fp16")];
            tensor<int32, [2]> var_1516_split_sizes_0 = const()[name = string("op_1516_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1516_axis_0 = const()[name = string("op_1516_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1516_0, tensor<fp16, [1, 8, 128, 128]> var_1516_1 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1516")];
            bool attn_weights_49_transpose_x_0 = const()[name = string("attn_weights_49_transpose_x_0"), val = bool(false)];
            bool attn_weights_49_transpose_y_0 = const()[name = string("attn_weights_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_0, transpose_y = attn_weights_49_transpose_y_0, x = var_1503_cast_fp16_0, y = var_1516_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_1519_to_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_53_cast_fp16 = add(x = attn_weights_51_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_53_cast_fp16")];
            int32 var_1523 = const()[name = string("op_1523"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_55_cast_fp16 = softmax(axis = var_1523, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1529_transpose_x_1 = const()[name = string("op_1529_transpose_x_1"), val = bool(true)];
            bool var_1529_transpose_y_1 = const()[name = string("op_1529_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1529_cast_fp16 = matmul(transpose_x = var_1529_transpose_x_1, transpose_y = var_1529_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1513_cast_fp16_0)[name = string("op_1529_cast_fp16")];
            bool attn_weights_57_transpose_x_0 = const()[name = string("attn_weights_57_transpose_x_0"), val = bool(false)];
            bool attn_weights_57_transpose_y_0 = const()[name = string("attn_weights_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_57_cast_fp16 = matmul(transpose_x = attn_weights_57_transpose_x_0, transpose_y = attn_weights_57_transpose_y_0, x = var_1503_cast_fp16_1, y = var_1516_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1531_to_fp16 = const()[name = string("op_1531_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_59_cast_fp16 = mul(x = attn_weights_57_cast_fp16, y = var_1531_to_fp16)[name = string("attn_weights_59_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_61_cast_fp16 = add(x = attn_weights_59_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_61_cast_fp16")];
            int32 var_1535 = const()[name = string("op_1535"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_63_cast_fp16 = softmax(axis = var_1535, x = attn_weights_61_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(true)];
            bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_63_cast_fp16, y = var_1513_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1543 = const()[name = string("op_1543"), val = int32(1)];
            bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_27_cast_fp16 = concat(axis = var_1543, interleave = attn_output_27_interleave_0, values = (var_1529_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1547_perm_0 = const()[name = string("op_1547_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_1547_cast_fp16 = transpose(perm = var_1547_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_195")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1547_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369560512)))];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1580_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1580_cast_fp16")];
            int32 var_1578 = const()[name = string("op_1578"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_29_cast_fp16 = concat(axis = var_1578, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1580_cast_fp16))[name = string("doubled_29_cast_fp16")];
            tensor<int32, [1]> out_15_axes_0 = const()[name = string("out_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_15_gamma_0_to_fp16 = const()[name = string("out_15_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377949184)))];
            fp16 var_1590_to_fp16 = const()[name = string("op_1590_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_1590_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1601_split_sizes_0 = const()[name = string("op_1601_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1601_axis_0 = const()[name = string("op_1601_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1601_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1601_cast_fp16_1 = split(axis = var_1601_axis_0, split_sizes = var_1601_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1601_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_1618_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1618_cast_fp16")];
            tensor<int32, [2]> var_1624_strides_0 = const()[name = string("op_1624_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1624_pad_type_0 = const()[name = string("op_1624_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1624_pad_0 = const()[name = string("op_1624_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1624_dilations_0 = const()[name = string("op_1624_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1624_groups_0 = const()[name = string("op_1624_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_1624_cast_fp16 = conv(dilations = var_1624_dilations_0, groups = var_1624_groups_0, pad = var_1624_pad_0, pad_type = var_1624_pad_type_0, strides = var_1624_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("op_1624_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_39_cast_fp16 = mul(x = var_1618_cast_fp16, y = var_1624_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1642_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1642_cast_fp16")];
            int32 var_1640 = const()[name = string("op_1640"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_33_cast_fp16 = concat(axis = var_1640, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1642_cast_fp16))[name = string("doubled_33_cast_fp16")];
            tensor<int32, [1]> out_17_axes_0 = const()[name = string("out_17_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_17_gamma_0_to_fp16 = const()[name = string("out_17_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377957440)))];
            fp16 var_1652_to_fp16 = const()[name = string("op_1652_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1652_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1663_split_sizes_0 = const()[name = string("op_1663_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1663_axis_0 = const()[name = string("op_1663_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1663_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1663_cast_fp16_1 = split(axis = var_1663_axis_0, split_sizes = var_1663_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1663_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377965696)))];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = var_1663_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_1720_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1720_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_1727_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1731_cast_fp16 = mul(x = x_41_cast_fp16, y = var_308_cast_fp16)[name = string("op_1731_cast_fp16")];
            tensor<int32, [2]> var_1732_split_sizes_0 = const()[name = string("op_1732_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1732_axis_0 = const()[name = string("op_1732_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_1732_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_1732_cast_fp16_1 = split(axis = var_1732_axis_0, split_sizes = var_1732_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1732_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_1734_cast_fp16 = mul(x = var_1732_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1734_cast_fp16")];
            int32 var_1736 = const()[name = string("op_1736"), val = int32(-2)];
            bool var_1737_interleave_0 = const()[name = string("op_1737_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_1737_cast_fp16 = concat(axis = var_1736, interleave = var_1737_interleave_0, values = (var_1734_cast_fp16, var_1732_cast_fp16_0))[name = string("op_1737_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1738_cast_fp16 = mul(x = var_1737_cast_fp16, y = var_315_cast_fp16)[name = string("op_1738_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_27_cast_fp16 = add(x = var_1731_cast_fp16, y = var_1738_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1744_cast_fp16 = mul(x = var_1720_cast_fp16, y = var_308_cast_fp16)[name = string("op_1744_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_1745_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = var_1720_cast_fp16)[name = string("op_1745_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_1747_cast_fp16 = mul(x = var_1745_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1747_cast_fp16")];
            int32 var_1749 = const()[name = string("op_1749"), val = int32(-2)];
            bool var_1750_interleave_0 = const()[name = string("op_1750_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_1750_cast_fp16 = concat(axis = var_1749, interleave = var_1750_interleave_0, values = (var_1747_cast_fp16, var_1745_cast_fp16_0))[name = string("op_1750_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1751_cast_fp16 = mul(x = var_1750_cast_fp16, y = var_315_cast_fp16)[name = string("op_1751_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_45_cast_fp16 = add(x = var_1744_cast_fp16, y = var_1751_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_194")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_114)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_116_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_116 = read_state(input = key_cache)[name = string("coreml_update_state_116")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1727_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_115)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_117_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_117 = read_state(input = value_cache)[name = string("coreml_update_state_117")];
            tensor<int32, [4]> var_1821_begin_0 = const()[name = string("op_1821_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1821_end_0 = const()[name = string("op_1821_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1821_end_mask_0 = const()[name = string("op_1821_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1821_cast_fp16 = slice_by_index(begin = var_1821_begin_0, end = var_1821_end_0, end_mask = var_1821_end_mask_0, x = coreml_update_state_116)[name = string("op_1821_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1824_axis_0 = const()[name = string("op_1824_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_1 = split(axis = var_1824_axis_0, split_sizes = tile_8, x = var_1821_cast_fp16)[name = string("op_1824_cast_fp16")];
            tensor<int32, [4]> var_1831_begin_0 = const()[name = string("op_1831_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1831_end_0 = const()[name = string("op_1831_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1831_end_mask_0 = const()[name = string("op_1831_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1831_cast_fp16 = slice_by_index(begin = var_1831_begin_0, end = var_1831_end_0, end_mask = var_1831_end_mask_0, x = coreml_update_state_117)[name = string("op_1831_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1834_axis_0 = const()[name = string("op_1834_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_1 = split(axis = var_1834_axis_0, split_sizes = tile_9, x = var_1831_cast_fp16)[name = string("op_1834_cast_fp16")];
            tensor<int32, [2]> var_1837_split_sizes_0 = const()[name = string("op_1837_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1837_axis_0 = const()[name = string("op_1837_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1837_0, tensor<fp16, [1, 8, 128, 128]> var_1837_1 = split(axis = var_1837_axis_0, split_sizes = var_1837_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1837")];
            bool attn_weights_65_transpose_x_0 = const()[name = string("attn_weights_65_transpose_x_0"), val = bool(false)];
            bool attn_weights_65_transpose_y_0 = const()[name = string("attn_weights_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_65_cast_fp16 = matmul(transpose_x = attn_weights_65_transpose_x_0, transpose_y = attn_weights_65_transpose_y_0, x = var_1824_cast_fp16_0, y = var_1837_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1840_to_fp16 = const()[name = string("op_1840_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_67_cast_fp16 = mul(x = attn_weights_65_cast_fp16, y = var_1840_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            int32 var_1844 = const()[name = string("op_1844"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_71_cast_fp16 = softmax(axis = var_1844, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1850_transpose_x_1 = const()[name = string("op_1850_transpose_x_1"), val = bool(true)];
            bool var_1850_transpose_y_1 = const()[name = string("op_1850_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1850_cast_fp16 = matmul(transpose_x = var_1850_transpose_x_1, transpose_y = var_1850_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1834_cast_fp16_0)[name = string("op_1850_cast_fp16")];
            bool attn_weights_73_transpose_x_0 = const()[name = string("attn_weights_73_transpose_x_0"), val = bool(false)];
            bool attn_weights_73_transpose_y_0 = const()[name = string("attn_weights_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_0, transpose_y = attn_weights_73_transpose_y_0, x = var_1824_cast_fp16_1, y = var_1837_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1852_to_fp16 = const()[name = string("op_1852_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1852_to_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_77_cast_fp16 = add(x = attn_weights_75_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_77_cast_fp16")];
            int32 var_1856 = const()[name = string("op_1856"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_79_cast_fp16 = softmax(axis = var_1856, x = attn_weights_77_cast_fp16)[name = string("attn_weights_79_cast_fp16")];
            bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(true)];
            bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_79_cast_fp16, y = var_1834_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1864 = const()[name = string("op_1864"), val = int32(1)];
            bool attn_output_35_interleave_0 = const()[name = string("attn_output_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_35_cast_fp16 = concat(axis = var_1864, interleave = attn_output_35_interleave_0, values = (var_1850_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1868_perm_0 = const()[name = string("op_1868_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_1868_cast_fp16 = transpose(perm = var_1868_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_192")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1868_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1901_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1901_cast_fp16")];
            int32 var_1899 = const()[name = string("op_1899"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_37_cast_fp16 = concat(axis = var_1899, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1901_cast_fp16))[name = string("doubled_37_cast_fp16")];
            tensor<int32, [1]> out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_19_gamma_0_to_fp16 = const()[name = string("out_19_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379014336)))];
            fp16 var_1911_to_fp16 = const()[name = string("op_1911_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1911_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1922_split_sizes_0 = const()[name = string("op_1922_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1922_axis_0 = const()[name = string("op_1922_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1922_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1922_cast_fp16_1 = split(axis = var_1922_axis_0, split_sizes = var_1922_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1922_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_1939_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<int32, [2]> var_1945_strides_0 = const()[name = string("op_1945_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1945_pad_type_0 = const()[name = string("op_1945_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1945_pad_0 = const()[name = string("op_1945_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1945_dilations_0 = const()[name = string("op_1945_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1945_groups_0 = const()[name = string("op_1945_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_1945_cast_fp16 = conv(dilations = var_1945_dilations_0, groups = var_1945_groups_0, pad = var_1945_pad_0, pad_type = var_1945_pad_type_0, strides = var_1945_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("op_1945_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_49_cast_fp16 = mul(x = var_1939_cast_fp16, y = var_1945_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_1963_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1963_cast_fp16")];
            int32 var_1961 = const()[name = string("op_1961"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_41_cast_fp16 = concat(axis = var_1961, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1963_cast_fp16))[name = string("doubled_41_cast_fp16")];
            tensor<int32, [1]> out_21_axes_0 = const()[name = string("out_21_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_21_gamma_0_to_fp16 = const()[name = string("out_21_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379022592)))];
            fp16 var_1973_to_fp16 = const()[name = string("op_1973_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1973_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_1984_split_sizes_0 = const()[name = string("op_1984_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1984_axis_0 = const()[name = string("op_1984_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1984_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1984_cast_fp16_1 = split(axis = var_1984_axis_0, split_sizes = var_1984_split_sizes_0, x = out_21_cast_fp16)[name = string("op_1984_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379030848)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_1984_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_2041_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2041_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_2048_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2048_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2052_cast_fp16 = mul(x = x_51_cast_fp16, y = var_308_cast_fp16)[name = string("op_2052_cast_fp16")];
            tensor<int32, [2]> var_2053_split_sizes_0 = const()[name = string("op_2053_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2053_axis_0 = const()[name = string("op_2053_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_2053_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_2053_cast_fp16_1 = split(axis = var_2053_axis_0, split_sizes = var_2053_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2053_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_2055_cast_fp16 = mul(x = var_2053_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2055_cast_fp16")];
            int32 var_2057 = const()[name = string("op_2057"), val = int32(-2)];
            bool var_2058_interleave_0 = const()[name = string("op_2058_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_2058_cast_fp16 = concat(axis = var_2057, interleave = var_2058_interleave_0, values = (var_2055_cast_fp16, var_2053_cast_fp16_0))[name = string("op_2058_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2059_cast_fp16 = mul(x = var_2058_cast_fp16, y = var_315_cast_fp16)[name = string("op_2059_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_33_cast_fp16 = add(x = var_2052_cast_fp16, y = var_2059_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2065_cast_fp16 = mul(x = var_2041_cast_fp16, y = var_308_cast_fp16)[name = string("op_2065_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_2066_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = var_2041_cast_fp16)[name = string("op_2066_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_2068_cast_fp16 = mul(x = var_2066_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2068_cast_fp16")];
            int32 var_2070 = const()[name = string("op_2070"), val = int32(-2)];
            bool var_2071_interleave_0 = const()[name = string("op_2071_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_2071_cast_fp16 = concat(axis = var_2070, interleave = var_2071_interleave_0, values = (var_2068_cast_fp16, var_2066_cast_fp16_0))[name = string("op_2071_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2072_cast_fp16 = mul(x = var_2071_cast_fp16, y = var_315_cast_fp16)[name = string("op_2072_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_55_cast_fp16 = add(x = var_2065_cast_fp16, y = var_2072_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_191")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_116)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_118_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_118 = read_state(input = key_cache)[name = string("coreml_update_state_118")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2048_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_117)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_119_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_119 = read_state(input = value_cache)[name = string("coreml_update_state_119")];
            tensor<int32, [4]> var_2142_begin_0 = const()[name = string("op_2142_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2142_end_0 = const()[name = string("op_2142_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2142_end_mask_0 = const()[name = string("op_2142_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2142_cast_fp16 = slice_by_index(begin = var_2142_begin_0, end = var_2142_end_0, end_mask = var_2142_end_mask_0, x = coreml_update_state_118)[name = string("op_2142_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2145_axis_0 = const()[name = string("op_2145_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_1 = split(axis = var_2145_axis_0, split_sizes = tile_10, x = var_2142_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [4]> var_2152_begin_0 = const()[name = string("op_2152_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2152_end_0 = const()[name = string("op_2152_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2152_end_mask_0 = const()[name = string("op_2152_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2152_cast_fp16 = slice_by_index(begin = var_2152_begin_0, end = var_2152_end_0, end_mask = var_2152_end_mask_0, x = coreml_update_state_119)[name = string("op_2152_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2155_axis_0 = const()[name = string("op_2155_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_1 = split(axis = var_2155_axis_0, split_sizes = tile_11, x = var_2152_cast_fp16)[name = string("op_2155_cast_fp16")];
            tensor<int32, [2]> var_2158_split_sizes_0 = const()[name = string("op_2158_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2158_axis_0 = const()[name = string("op_2158_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_2158_0, tensor<fp16, [1, 8, 128, 128]> var_2158_1 = split(axis = var_2158_axis_0, split_sizes = var_2158_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2158")];
            bool attn_weights_81_transpose_x_0 = const()[name = string("attn_weights_81_transpose_x_0"), val = bool(false)];
            bool attn_weights_81_transpose_y_0 = const()[name = string("attn_weights_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_81_cast_fp16 = matmul(transpose_x = attn_weights_81_transpose_x_0, transpose_y = attn_weights_81_transpose_y_0, x = var_2145_cast_fp16_0, y = var_2158_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2161_to_fp16 = const()[name = string("op_2161_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_83_cast_fp16 = mul(x = attn_weights_81_cast_fp16, y = var_2161_to_fp16)[name = string("attn_weights_83_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_85_cast_fp16 = add(x = attn_weights_83_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_85_cast_fp16")];
            int32 var_2165 = const()[name = string("op_2165"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_87_cast_fp16 = softmax(axis = var_2165, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2171_transpose_x_1 = const()[name = string("op_2171_transpose_x_1"), val = bool(true)];
            bool var_2171_transpose_y_1 = const()[name = string("op_2171_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2171_cast_fp16 = matmul(transpose_x = var_2171_transpose_x_1, transpose_y = var_2171_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2155_cast_fp16_0)[name = string("op_2171_cast_fp16")];
            bool attn_weights_89_transpose_x_0 = const()[name = string("attn_weights_89_transpose_x_0"), val = bool(false)];
            bool attn_weights_89_transpose_y_0 = const()[name = string("attn_weights_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_89_cast_fp16 = matmul(transpose_x = attn_weights_89_transpose_x_0, transpose_y = attn_weights_89_transpose_y_0, x = var_2145_cast_fp16_1, y = var_2158_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2173_to_fp16 = const()[name = string("op_2173_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_91_cast_fp16 = mul(x = attn_weights_89_cast_fp16, y = var_2173_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            int32 var_2177 = const()[name = string("op_2177"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_95_cast_fp16 = softmax(axis = var_2177, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")];
            bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(true)];
            bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_95_cast_fp16, y = var_2155_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2185 = const()[name = string("op_2185"), val = int32(1)];
            bool attn_output_43_interleave_0 = const()[name = string("attn_output_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_43_cast_fp16 = concat(axis = var_2185, interleave = attn_output_43_interleave_0, values = (var_2171_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2189_perm_0 = const()[name = string("op_2189_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_2189_cast_fp16 = transpose(perm = var_2189_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_189")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2189_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_2222_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2222_cast_fp16")];
            int32 var_2220 = const()[name = string("op_2220"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_45_cast_fp16 = concat(axis = var_2220, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2222_cast_fp16))[name = string("doubled_45_cast_fp16")];
            tensor<int32, [1]> out_23_axes_0 = const()[name = string("out_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_23_gamma_0_to_fp16 = const()[name = string("out_23_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380079488)))];
            fp16 var_2232_to_fp16 = const()[name = string("op_2232_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_2232_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2243_split_sizes_0 = const()[name = string("op_2243_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2243_axis_0 = const()[name = string("op_2243_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2243_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2243_cast_fp16_1 = split(axis = var_2243_axis_0, split_sizes = var_2243_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2243_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_2260_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<int32, [2]> var_2266_strides_0 = const()[name = string("op_2266_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2266_pad_type_0 = const()[name = string("op_2266_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2266_pad_0 = const()[name = string("op_2266_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2266_dilations_0 = const()[name = string("op_2266_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2266_groups_0 = const()[name = string("op_2266_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_2266_cast_fp16 = conv(dilations = var_2266_dilations_0, groups = var_2266_groups_0, pad = var_2266_pad_0, pad_type = var_2266_pad_type_0, strides = var_2266_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("op_2266_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_59_cast_fp16 = mul(x = var_2260_cast_fp16, y = var_2266_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_2284_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2284_cast_fp16")];
            int32 var_2282 = const()[name = string("op_2282"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_49_cast_fp16 = concat(axis = var_2282, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2284_cast_fp16))[name = string("doubled_49_cast_fp16")];
            tensor<int32, [1]> out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_25_gamma_0_to_fp16 = const()[name = string("out_25_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380087744)))];
            fp16 var_2294_to_fp16 = const()[name = string("op_2294_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_2294_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2305_split_sizes_0 = const()[name = string("op_2305_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2305_axis_0 = const()[name = string("op_2305_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2305_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2305_cast_fp16_1 = split(axis = var_2305_axis_0, split_sizes = var_2305_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380096000)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2305_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_2362_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2362_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_2369_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2369_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2373_cast_fp16 = mul(x = x_61_cast_fp16, y = var_308_cast_fp16)[name = string("op_2373_cast_fp16")];
            tensor<int32, [2]> var_2374_split_sizes_0 = const()[name = string("op_2374_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2374_axis_0 = const()[name = string("op_2374_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_2374_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_2374_cast_fp16_1 = split(axis = var_2374_axis_0, split_sizes = var_2374_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2376_cast_fp16")];
            int32 var_2378 = const()[name = string("op_2378"), val = int32(-2)];
            bool var_2379_interleave_0 = const()[name = string("op_2379_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_2379_cast_fp16 = concat(axis = var_2378, interleave = var_2379_interleave_0, values = (var_2376_cast_fp16, var_2374_cast_fp16_0))[name = string("op_2379_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2380_cast_fp16 = mul(x = var_2379_cast_fp16, y = var_315_cast_fp16)[name = string("op_2380_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_39_cast_fp16 = add(x = var_2373_cast_fp16, y = var_2380_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2386_cast_fp16 = mul(x = var_2362_cast_fp16, y = var_308_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_2387_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = var_2362_cast_fp16)[name = string("op_2387_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_2389_cast_fp16 = mul(x = var_2387_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2389_cast_fp16")];
            int32 var_2391 = const()[name = string("op_2391"), val = int32(-2)];
            bool var_2392_interleave_0 = const()[name = string("op_2392_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_2392_cast_fp16 = concat(axis = var_2391, interleave = var_2392_interleave_0, values = (var_2389_cast_fp16, var_2387_cast_fp16_0))[name = string("op_2392_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2393_cast_fp16 = mul(x = var_2392_cast_fp16, y = var_315_cast_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_65_cast_fp16 = add(x = var_2386_cast_fp16, y = var_2393_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_118)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_120_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_120 = read_state(input = key_cache)[name = string("coreml_update_state_120")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2369_cast_fp16)[name = string("transpose_187")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_119)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_121_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_121 = read_state(input = value_cache)[name = string("coreml_update_state_121")];
            tensor<int32, [4]> var_2463_begin_0 = const()[name = string("op_2463_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2463_end_0 = const()[name = string("op_2463_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2463_end_mask_0 = const()[name = string("op_2463_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2463_cast_fp16 = slice_by_index(begin = var_2463_begin_0, end = var_2463_end_0, end_mask = var_2463_end_mask_0, x = coreml_update_state_120)[name = string("op_2463_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2466_axis_0 = const()[name = string("op_2466_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_1 = split(axis = var_2466_axis_0, split_sizes = tile_12, x = var_2463_cast_fp16)[name = string("op_2466_cast_fp16")];
            tensor<int32, [4]> var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, x = coreml_update_state_121)[name = string("op_2473_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2476_axis_0 = const()[name = string("op_2476_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_1 = split(axis = var_2476_axis_0, split_sizes = tile_13, x = var_2473_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<int32, [2]> var_2479_split_sizes_0 = const()[name = string("op_2479_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2479_axis_0 = const()[name = string("op_2479_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_2479_0, tensor<fp16, [1, 8, 128, 128]> var_2479_1 = split(axis = var_2479_axis_0, split_sizes = var_2479_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2479")];
            bool attn_weights_97_transpose_x_0 = const()[name = string("attn_weights_97_transpose_x_0"), val = bool(false)];
            bool attn_weights_97_transpose_y_0 = const()[name = string("attn_weights_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_0, transpose_y = attn_weights_97_transpose_y_0, x = var_2466_cast_fp16_0, y = var_2479_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2482_to_fp16 = const()[name = string("op_2482_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_2482_to_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_101_cast_fp16 = add(x = attn_weights_99_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_101_cast_fp16")];
            int32 var_2486 = const()[name = string("op_2486"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_103_cast_fp16 = softmax(axis = var_2486, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2492_transpose_x_1 = const()[name = string("op_2492_transpose_x_1"), val = bool(true)];
            bool var_2492_transpose_y_1 = const()[name = string("op_2492_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2492_cast_fp16 = matmul(transpose_x = var_2492_transpose_x_1, transpose_y = var_2492_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2476_cast_fp16_0)[name = string("op_2492_cast_fp16")];
            bool attn_weights_105_transpose_x_0 = const()[name = string("attn_weights_105_transpose_x_0"), val = bool(false)];
            bool attn_weights_105_transpose_y_0 = const()[name = string("attn_weights_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_105_cast_fp16 = matmul(transpose_x = attn_weights_105_transpose_x_0, transpose_y = attn_weights_105_transpose_y_0, x = var_2466_cast_fp16_1, y = var_2479_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2494_to_fp16 = const()[name = string("op_2494_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_107_cast_fp16 = mul(x = attn_weights_105_cast_fp16, y = var_2494_to_fp16)[name = string("attn_weights_107_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_109_cast_fp16 = add(x = attn_weights_107_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_109_cast_fp16")];
            int32 var_2498 = const()[name = string("op_2498"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_111_cast_fp16 = softmax(axis = var_2498, x = attn_weights_109_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(true)];
            bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_111_cast_fp16, y = var_2476_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2506 = const()[name = string("op_2506"), val = int32(1)];
            bool attn_output_51_interleave_0 = const()[name = string("attn_output_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_51_cast_fp16 = concat(axis = var_2506, interleave = attn_output_51_interleave_0, values = (var_2492_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2510_perm_0 = const()[name = string("op_2510_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_2510_cast_fp16 = transpose(perm = var_2510_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_186")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2510_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_2543_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2543_cast_fp16")];
            int32 var_2541 = const()[name = string("op_2541"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_53_cast_fp16 = concat(axis = var_2541, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2543_cast_fp16))[name = string("doubled_53_cast_fp16")];
            tensor<int32, [1]> out_27_axes_0 = const()[name = string("out_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_27_gamma_0_to_fp16 = const()[name = string("out_27_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381144640)))];
            fp16 var_2553_to_fp16 = const()[name = string("op_2553_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_2553_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2564_split_sizes_0 = const()[name = string("op_2564_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2564_axis_0 = const()[name = string("op_2564_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2564_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2564_cast_fp16_1 = split(axis = var_2564_axis_0, split_sizes = var_2564_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_2581_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [2]> var_2587_strides_0 = const()[name = string("op_2587_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2587_pad_type_0 = const()[name = string("op_2587_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2587_pad_0 = const()[name = string("op_2587_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2587_dilations_0 = const()[name = string("op_2587_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2587_groups_0 = const()[name = string("op_2587_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_2587_cast_fp16 = conv(dilations = var_2587_dilations_0, groups = var_2587_groups_0, pad = var_2587_pad_0, pad_type = var_2587_pad_type_0, strides = var_2587_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("op_2587_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_69_cast_fp16 = mul(x = var_2581_cast_fp16, y = var_2587_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381152896)))];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_to_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_2605_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2605_cast_fp16")];
            int32 var_2603 = const()[name = string("op_2603"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_57_cast_fp16 = concat(axis = var_2603, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2605_cast_fp16))[name = string("doubled_57_cast_fp16")];
            tensor<int32, [1]> out_29_axes_0 = const()[name = string("out_29_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_29_gamma_0_to_fp16 = const()[name = string("out_29_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406318784)))];
            fp16 var_2615_to_fp16 = const()[name = string("op_2615_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_2615_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2626_split_sizes_0 = const()[name = string("op_2626_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2626_axis_0 = const()[name = string("op_2626_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2626_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2626_cast_fp16_1 = split(axis = var_2626_axis_0, split_sizes = var_2626_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2626_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406327040)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2626_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_2683_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_2690_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2690_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2694_cast_fp16 = mul(x = x_71_cast_fp16, y = var_308_cast_fp16)[name = string("op_2694_cast_fp16")];
            tensor<int32, [2]> var_2695_split_sizes_0 = const()[name = string("op_2695_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2695_axis_0 = const()[name = string("op_2695_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_2695_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_2695_cast_fp16_1 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2695_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_2697_cast_fp16 = mul(x = var_2695_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2697_cast_fp16")];
            int32 var_2699 = const()[name = string("op_2699"), val = int32(-2)];
            bool var_2700_interleave_0 = const()[name = string("op_2700_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_2700_cast_fp16 = concat(axis = var_2699, interleave = var_2700_interleave_0, values = (var_2697_cast_fp16, var_2695_cast_fp16_0))[name = string("op_2700_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2701_cast_fp16 = mul(x = var_2700_cast_fp16, y = var_315_cast_fp16)[name = string("op_2701_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_45_cast_fp16 = add(x = var_2694_cast_fp16, y = var_2701_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2707_cast_fp16 = mul(x = var_2683_cast_fp16, y = var_308_cast_fp16)[name = string("op_2707_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_2708_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = var_2683_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_2710_cast_fp16 = mul(x = var_2708_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2710_cast_fp16")];
            int32 var_2712 = const()[name = string("op_2712"), val = int32(-2)];
            bool var_2713_interleave_0 = const()[name = string("op_2713_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_2713_cast_fp16 = concat(axis = var_2712, interleave = var_2713_interleave_0, values = (var_2710_cast_fp16, var_2708_cast_fp16_0))[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2714_cast_fp16 = mul(x = var_2713_cast_fp16, y = var_315_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_75_cast_fp16 = add(x = var_2707_cast_fp16, y = var_2714_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_185")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_120)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_122_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_122 = read_state(input = key_cache)[name = string("coreml_update_state_122")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2690_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_121)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_123_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_123 = read_state(input = value_cache)[name = string("coreml_update_state_123")];
            tensor<int32, [4]> var_2784_begin_0 = const()[name = string("op_2784_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2784_end_0 = const()[name = string("op_2784_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2784_end_mask_0 = const()[name = string("op_2784_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2784_cast_fp16 = slice_by_index(begin = var_2784_begin_0, end = var_2784_end_0, end_mask = var_2784_end_mask_0, x = coreml_update_state_122)[name = string("op_2784_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2787_axis_0 = const()[name = string("op_2787_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_1 = split(axis = var_2787_axis_0, split_sizes = tile_14, x = var_2784_cast_fp16)[name = string("op_2787_cast_fp16")];
            tensor<int32, [4]> var_2794_begin_0 = const()[name = string("op_2794_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2794_end_0 = const()[name = string("op_2794_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2794_end_mask_0 = const()[name = string("op_2794_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2794_cast_fp16 = slice_by_index(begin = var_2794_begin_0, end = var_2794_end_0, end_mask = var_2794_end_mask_0, x = coreml_update_state_123)[name = string("op_2794_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2797_axis_0 = const()[name = string("op_2797_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_1 = split(axis = var_2797_axis_0, split_sizes = tile_15, x = var_2794_cast_fp16)[name = string("op_2797_cast_fp16")];
            tensor<int32, [2]> var_2800_split_sizes_0 = const()[name = string("op_2800_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2800_axis_0 = const()[name = string("op_2800_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_2800_0, tensor<fp16, [1, 8, 128, 128]> var_2800_1 = split(axis = var_2800_axis_0, split_sizes = var_2800_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2800")];
            bool attn_weights_113_transpose_x_0 = const()[name = string("attn_weights_113_transpose_x_0"), val = bool(false)];
            bool attn_weights_113_transpose_y_0 = const()[name = string("attn_weights_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_113_cast_fp16 = matmul(transpose_x = attn_weights_113_transpose_x_0, transpose_y = attn_weights_113_transpose_y_0, x = var_2787_cast_fp16_0, y = var_2800_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2803_to_fp16 = const()[name = string("op_2803_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_115_cast_fp16 = mul(x = attn_weights_113_cast_fp16, y = var_2803_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            int32 var_2807 = const()[name = string("op_2807"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_119_cast_fp16 = softmax(axis = var_2807, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2813_transpose_x_1 = const()[name = string("op_2813_transpose_x_1"), val = bool(true)];
            bool var_2813_transpose_y_1 = const()[name = string("op_2813_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2813_cast_fp16 = matmul(transpose_x = var_2813_transpose_x_1, transpose_y = var_2813_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2797_cast_fp16_0)[name = string("op_2813_cast_fp16")];
            bool attn_weights_121_transpose_x_0 = const()[name = string("attn_weights_121_transpose_x_0"), val = bool(false)];
            bool attn_weights_121_transpose_y_0 = const()[name = string("attn_weights_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_0, transpose_y = attn_weights_121_transpose_y_0, x = var_2787_cast_fp16_1, y = var_2800_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2815_to_fp16 = const()[name = string("op_2815_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_2815_to_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_125_cast_fp16 = add(x = attn_weights_123_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_125_cast_fp16")];
            int32 var_2819 = const()[name = string("op_2819"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_127_cast_fp16 = softmax(axis = var_2819, x = attn_weights_125_cast_fp16)[name = string("attn_weights_127_cast_fp16")];
            bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(true)];
            bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_127_cast_fp16, y = var_2797_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2827 = const()[name = string("op_2827"), val = int32(1)];
            bool attn_output_59_interleave_0 = const()[name = string("attn_output_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_59_cast_fp16 = concat(axis = var_2827, interleave = attn_output_59_interleave_0, values = (var_2813_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2831_perm_0 = const()[name = string("op_2831_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_2831_cast_fp16 = transpose(perm = var_2831_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_183")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2831_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_2864_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2864_cast_fp16")];
            int32 var_2862 = const()[name = string("op_2862"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_61_cast_fp16 = concat(axis = var_2862, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2864_cast_fp16))[name = string("doubled_61_cast_fp16")];
            tensor<int32, [1]> out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_31_gamma_0_to_fp16 = const()[name = string("out_31_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407375680)))];
            fp16 var_2874_to_fp16 = const()[name = string("op_2874_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_2874_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2885_split_sizes_0 = const()[name = string("op_2885_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2885_axis_0 = const()[name = string("op_2885_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2885_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2885_cast_fp16_1 = split(axis = var_2885_axis_0, split_sizes = var_2885_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2885_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2885_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_2902_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2902_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407383936)))];
            tensor<int32, [2]> var_2908_strides_0 = const()[name = string("op_2908_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2908_pad_type_0 = const()[name = string("op_2908_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2908_pad_0 = const()[name = string("op_2908_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2908_dilations_0 = const()[name = string("op_2908_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2908_groups_0 = const()[name = string("op_2908_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_2908_cast_fp16 = conv(dilations = var_2908_dilations_0, groups = var_2908_groups_0, pad = var_2908_pad_0, pad_type = var_2908_pad_type_0, strides = var_2908_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2885_cast_fp16_0)[name = string("op_2908_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_79_cast_fp16 = mul(x = var_2902_cast_fp16, y = var_2908_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432549824)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_2926_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2926_cast_fp16")];
            int32 var_2924 = const()[name = string("op_2924"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_65_cast_fp16 = concat(axis = var_2924, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2926_cast_fp16))[name = string("doubled_65_cast_fp16")];
            tensor<int32, [1]> out_33_axes_0 = const()[name = string("out_33_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_33_gamma_0_to_fp16 = const()[name = string("out_33_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457715712)))];
            fp16 var_2936_to_fp16 = const()[name = string("op_2936_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_2936_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2947_split_sizes_0 = const()[name = string("op_2947_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2947_axis_0 = const()[name = string("op_2947_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2947_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2947_cast_fp16_1 = split(axis = var_2947_axis_0, split_sizes = var_2947_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2947_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457723968)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466112640)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467161280)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_3004_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3004_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_3011_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3011_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3015_cast_fp16 = mul(x = x_81_cast_fp16, y = var_308_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<int32, [2]> var_3016_split_sizes_0 = const()[name = string("op_3016_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3016_axis_0 = const()[name = string("op_3016_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_3016_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_3016_cast_fp16_1 = split(axis = var_3016_axis_0, split_sizes = var_3016_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3016_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_3018_cast_fp16 = mul(x = var_3016_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3018_cast_fp16")];
            int32 var_3020 = const()[name = string("op_3020"), val = int32(-2)];
            bool var_3021_interleave_0 = const()[name = string("op_3021_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_3021_cast_fp16 = concat(axis = var_3020, interleave = var_3021_interleave_0, values = (var_3018_cast_fp16, var_3016_cast_fp16_0))[name = string("op_3021_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3022_cast_fp16 = mul(x = var_3021_cast_fp16, y = var_315_cast_fp16)[name = string("op_3022_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_51_cast_fp16 = add(x = var_3015_cast_fp16, y = var_3022_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3028_cast_fp16 = mul(x = var_3004_cast_fp16, y = var_308_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_3029_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = var_3004_cast_fp16)[name = string("op_3029_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_3031_cast_fp16 = mul(x = var_3029_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3031_cast_fp16")];
            int32 var_3033 = const()[name = string("op_3033"), val = int32(-2)];
            bool var_3034_interleave_0 = const()[name = string("op_3034_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_3034_cast_fp16 = concat(axis = var_3033, interleave = var_3034_interleave_0, values = (var_3031_cast_fp16, var_3029_cast_fp16_0))[name = string("op_3034_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3035_cast_fp16 = mul(x = var_3034_cast_fp16, y = var_315_cast_fp16)[name = string("op_3035_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_85_cast_fp16 = add(x = var_3028_cast_fp16, y = var_3035_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_182")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_122)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_124_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_124 = read_state(input = key_cache)[name = string("coreml_update_state_124")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3011_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_123)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_125_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_125 = read_state(input = value_cache)[name = string("coreml_update_state_125")];
            tensor<int32, [4]> var_3105_begin_0 = const()[name = string("op_3105_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3105_end_0 = const()[name = string("op_3105_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3105_end_mask_0 = const()[name = string("op_3105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3105_cast_fp16 = slice_by_index(begin = var_3105_begin_0, end = var_3105_end_0, end_mask = var_3105_end_mask_0, x = coreml_update_state_124)[name = string("op_3105_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3108_axis_0 = const()[name = string("op_3108_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_1 = split(axis = var_3108_axis_0, split_sizes = tile_16, x = var_3105_cast_fp16)[name = string("op_3108_cast_fp16")];
            tensor<int32, [4]> var_3115_begin_0 = const()[name = string("op_3115_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3115_end_0 = const()[name = string("op_3115_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3115_end_mask_0 = const()[name = string("op_3115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3115_cast_fp16 = slice_by_index(begin = var_3115_begin_0, end = var_3115_end_0, end_mask = var_3115_end_mask_0, x = coreml_update_state_125)[name = string("op_3115_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3118_axis_0 = const()[name = string("op_3118_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_1 = split(axis = var_3118_axis_0, split_sizes = tile_17, x = var_3115_cast_fp16)[name = string("op_3118_cast_fp16")];
            tensor<int32, [2]> var_3121_split_sizes_0 = const()[name = string("op_3121_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3121_axis_0 = const()[name = string("op_3121_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_3121_0, tensor<fp16, [1, 8, 128, 128]> var_3121_1 = split(axis = var_3121_axis_0, split_sizes = var_3121_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3121")];
            bool attn_weights_129_transpose_x_0 = const()[name = string("attn_weights_129_transpose_x_0"), val = bool(false)];
            bool attn_weights_129_transpose_y_0 = const()[name = string("attn_weights_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_129_cast_fp16 = matmul(transpose_x = attn_weights_129_transpose_x_0, transpose_y = attn_weights_129_transpose_y_0, x = var_3108_cast_fp16_0, y = var_3121_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3124_to_fp16 = const()[name = string("op_3124_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_131_cast_fp16 = mul(x = attn_weights_129_cast_fp16, y = var_3124_to_fp16)[name = string("attn_weights_131_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_133_cast_fp16 = add(x = attn_weights_131_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_133_cast_fp16")];
            int32 var_3128 = const()[name = string("op_3128"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_135_cast_fp16 = softmax(axis = var_3128, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3134_transpose_x_1 = const()[name = string("op_3134_transpose_x_1"), val = bool(true)];
            bool var_3134_transpose_y_1 = const()[name = string("op_3134_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_3134_cast_fp16 = matmul(transpose_x = var_3134_transpose_x_1, transpose_y = var_3134_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3118_cast_fp16_0)[name = string("op_3134_cast_fp16")];
            bool attn_weights_137_transpose_x_0 = const()[name = string("attn_weights_137_transpose_x_0"), val = bool(false)];
            bool attn_weights_137_transpose_y_0 = const()[name = string("attn_weights_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_137_cast_fp16 = matmul(transpose_x = attn_weights_137_transpose_x_0, transpose_y = attn_weights_137_transpose_y_0, x = var_3108_cast_fp16_1, y = var_3121_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3136_to_fp16 = const()[name = string("op_3136_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_139_cast_fp16 = mul(x = attn_weights_137_cast_fp16, y = var_3136_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            int32 var_3140 = const()[name = string("op_3140"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_cast_fp16 = softmax(axis = var_3140, x = attn_weights_141_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(true)];
            bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_cast_fp16, y = var_3118_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3148 = const()[name = string("op_3148"), val = int32(1)];
            bool attn_output_67_interleave_0 = const()[name = string("attn_output_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_67_cast_fp16 = concat(axis = var_3148, interleave = attn_output_67_interleave_0, values = (var_3134_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3152_perm_0 = const()[name = string("op_3152_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_3152_cast_fp16 = transpose(perm = var_3152_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_cast_fp16 = reshape(shape = concat_107x, x = var_3152_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468209920)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_3185_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3185_cast_fp16")];
            int32 var_3183 = const()[name = string("op_3183"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_69_cast_fp16 = concat(axis = var_3183, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3185_cast_fp16))[name = string("doubled_69_cast_fp16")];
            tensor<int32, [1]> out_35_axes_0 = const()[name = string("out_35_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_35_gamma_0_to_fp16 = const()[name = string("out_35_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476598592)))];
            fp16 var_3195_to_fp16 = const()[name = string("op_3195_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_3195_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3206_split_sizes_0 = const()[name = string("op_3206_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3206_axis_0 = const()[name = string("op_3206_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3206_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3206_cast_fp16_1 = split(axis = var_3206_axis_0, split_sizes = var_3206_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3206_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476606848)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_8_mlp_gate_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_3223_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3223_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501772736)))];
            tensor<int32, [2]> var_3229_strides_0 = const()[name = string("op_3229_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3229_pad_type_0 = const()[name = string("op_3229_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3229_pad_0 = const()[name = string("op_3229_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3229_dilations_0 = const()[name = string("op_3229_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3229_groups_0 = const()[name = string("op_3229_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_3229_cast_fp16 = conv(dilations = var_3229_dilations_0, groups = var_3229_groups_0, pad = var_3229_pad_0, pad_type = var_3229_pad_type_0, strides = var_3229_strides_0, weight = layers_8_mlp_up_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("op_3229_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_cast_fp16 = mul(x = var_3223_cast_fp16, y = var_3229_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526938624)))];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_3247_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3247_cast_fp16")];
            int32 var_3245 = const()[name = string("op_3245"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_73_cast_fp16 = concat(axis = var_3245, interleave = doubled_73_interleave_0, values = (hidden_states_cast_fp16, var_3247_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_axes_0 = const()[name = string("out_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_gamma_0_to_fp16 = const()[name = string("out_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552104512)))];
            fp16 var_3257_to_fp16 = const()[name = string("op_3257_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3257_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3268_split_sizes_0 = const()[name = string("op_3268_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3268_axis_0 = const()[name = string("op_3268_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states, tensor<fp16, [1, 2048, 1, 128]> var_3268_cast_fp16_1 = split(axis = var_3268_axis_0, split_sizes = var_3268_split_sizes_0, x = out_cast_fp16)[name = string("op_3268_cast_fp16")];
        } -> (hidden_states);
    func length_16<ios18>(tensor<fp32, [1, 2048, 1, 16]> inputs_embeds, state<tensor<fp16, [9, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [16]> position_index_seed, state<tensor<fp16, [9, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725952))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17321280))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17308928))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17327488))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29922816))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29910464))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29929024))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42516160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42512000))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42518272))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243840))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243264))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47244160))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51442688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51438528))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444800))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64040128))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64027776))))[name = string("layers_1_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64046336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76633472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76629312))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76635584))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80834112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80829952))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80836224))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361152))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81360576))))[name = string("layers_2_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361472))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85560000))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85555840))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85562112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98145088))))[name = string("layers_2_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110758976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110746624))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110765184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123352320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123348160))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123354432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127552960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127548800))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127555072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128079424))))[name = string("layers_3_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080320))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140675648))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140663296))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140681856))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153277184))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153264832))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153283392))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165870528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165866368))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165872640))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170071168))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170067008))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170073280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170597632))))[name = string("layers_4_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174797056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174792896))))[name = string("layers_4_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174799168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187394496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187382144))))[name = string("layers_4_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187400704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199996032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199983680))))[name = string("layers_4_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_4_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200002240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212589376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212585216))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212591488))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216790016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216785856))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216792128))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317056))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217316480))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317376))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221515904))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221511744))))[name = string("layers_5_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221518016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234113344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234100992))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234119552))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246714880))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246702528))))[name = string("layers_5_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_5_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246721088))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259308224))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259304064))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259310336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263508864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263504704))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263510976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035328))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264036224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268234752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268230592))))[name = string("layers_6_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268236864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280832192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280819840))))[name = string("layers_6_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280838400))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293433728))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293421376))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293439936))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297638464))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297634304))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297640576))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165504))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298164928))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165824))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302364352))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302360192))))[name = string("layers_7_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302366464))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314961792))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314949440))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(16)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [16]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_280 = const()[name = string("op_280"), val = int32(0)];
            bool var_282_exclusive_0 = const()[name = string("op_282_exclusive_0"), val = bool(false)];
            bool var_282_reverse_0 = const()[name = string("op_282_reverse_0"), val = bool(false)];
            tensor<fp16, [16]> var_282_cast_fp16 = cumsum(axis = var_280, exclusive = var_282_exclusive_0, reverse = var_282_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_282_cast_fp16")];
            fp16 var_284_promoted_to_fp16 = const()[name = string("op_284_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [16]> position_offsets_cast_fp16 = sub(x = var_282_cast_fp16, y = var_284_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 16]> var_287_cast_fp16 = expand_dims(axes = var_287_axes_0, x = position_offsets_cast_fp16)[name = string("op_287_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_15")];
            tensor<fp16, [1, 16]> position_ids_1_cast_fp16 = add(x = var_287_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 16]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_14")];
            tensor<bool, [1, 16]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 16]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 16]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314968000)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_6_y_0 = const()[name = string("greater_equal_6_y_0"), val = int32(0)];
            tensor<bool, [1, 16]> greater_equal_6 = greater_equal(x = select_0, y = greater_equal_6_y_0)[name = string("greater_equal_6")];
            int32 slice_by_index_6 = const()[name = string("slice_by_index_6"), val = int32(32768)];
            tensor<int32, [1, 16]> add_6 = add(x = select_0, y = slice_by_index_6)[name = string("add_6")];
            tensor<int32, [1, 16]> select_6 = select(a = select_0, b = add_6, cond = greater_equal_6)[name = string("select_6")];
            int32 cos_1_cast_fp16_axis_3 = const()[name = string("cos_1_cast_fp16_axis_3"), val = int32(0)];
            tensor<fp16, [1, 16, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_3, batch_dims = cos_1_batch_dims_0, indices = select_6, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323356672)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_3 = const()[name = string("sin_1_cast_fp16_axis_3"), val = int32(0)];
            tensor<fp16, [1, 16, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_3, batch_dims = sin_1_batch_dims_0, indices = select_6, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_306_perm_0 = const()[name = string("op_306_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_308_axes_0 = const()[name = string("op_308_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 16]> var_306_cast_fp16 = transpose(perm = var_306_perm_0, x = cos_1_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 1, 128, 16]> var_308_cast_fp16 = expand_dims(axes = var_308_axes_0, x = var_306_cast_fp16)[name = string("op_308_cast_fp16")];
            tensor<int32, [3]> var_313_perm_0 = const()[name = string("op_313_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 16]> var_313_cast_fp16 = transpose(perm = var_313_perm_0, x = sin_1_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 1, 128, 16]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = var_313_cast_fp16)[name = string("op_315_cast_fp16")];
            tensor<int32, [1]> var_334_axes_0 = const()[name = string("op_334_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 16, 1]> var_334 = expand_dims(axes = var_334_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_334")];
            tensor<int32, [1, 1, 2048]> var_327 = const()[name = string("op_327"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331745344)))];
            tensor<bool, [1, 16, 2048]> var_335 = greater(x = var_327, y = var_334)[name = string("op_335")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_342_axes_0 = const()[name = string("op_342_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 16, 2048]> var_335_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_335)[name = string("cast_13")];
            tensor<fp16, [1, 1, 16, 2048]> var_342_cast_fp16 = expand_dims(axes = var_342_axes_0, x = var_335_to_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_346_promoted_to_fp16 = const()[name = string("op_346_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 16]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_342_cast_fp16)[name = string("transpose_117")];
            tensor<bool, [1, 1, 2048, 16]> var_347_cast_fp16 = equal(x = mask_cast_fp16, y = var_346_promoted_to_fp16)[name = string("op_347_cast_fp16")];
            fp16 var_348_to_fp16 = const()[name = string("op_348_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 16]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_348_to_fp16, cond = var_347_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_12")];
            tensor<fp16, [1, 2048, 1, 16]> var_358_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_358_cast_fp16")];
            int32 var_356 = const()[name = string("op_356"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_1_cast_fp16 = concat(axis = var_356, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_358_cast_fp16))[name = string("doubled_1_cast_fp16")];
            tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_1_gamma_0_to_fp16 = const()[name = string("out_1_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331753600)))];
            fp16 var_368_to_fp16 = const()[name = string("op_368_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_368_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_379_split_sizes_0 = const()[name = string("op_379_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_379_axis_0 = const()[name = string("op_379_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_379_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_379_cast_fp16_1 = split(axis = var_379_axis_0, split_sizes = var_379_split_sizes_0, x = out_1_cast_fp16)[name = string("op_379_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331761856)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_379_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_436_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_436_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_443_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_443_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_447_cast_fp16 = mul(x = x_1_cast_fp16, y = var_308_cast_fp16)[name = string("op_447_cast_fp16")];
            tensor<int32, [2]> var_448_split_sizes_0 = const()[name = string("op_448_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_448_axis_0 = const()[name = string("op_448_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_448_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_448_cast_fp16_1 = split(axis = var_448_axis_0, split_sizes = var_448_split_sizes_0, x = x_1_cast_fp16)[name = string("op_448_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_450_cast_fp16 = mul(x = var_448_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_450_cast_fp16")];
            int32 var_452 = const()[name = string("op_452"), val = int32(-2)];
            bool var_453_interleave_0 = const()[name = string("op_453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_453_cast_fp16 = concat(axis = var_452, interleave = var_453_interleave_0, values = (var_450_cast_fp16, var_448_cast_fp16_0))[name = string("op_453_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_454_cast_fp16 = mul(x = var_453_cast_fp16, y = var_315_cast_fp16)[name = string("op_454_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_3_cast_fp16 = add(x = var_447_cast_fp16, y = var_454_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_460_cast_fp16 = mul(x = var_436_cast_fp16, y = var_308_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [2]> var_461_split_sizes_0 = const()[name = string("op_461_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_461_axis_0 = const()[name = string("op_461_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_461_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_461_cast_fp16_1 = split(axis = var_461_axis_0, split_sizes = var_461_split_sizes_0, x = var_436_cast_fp16)[name = string("op_461_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_463_cast_fp16 = mul(x = var_461_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_463_cast_fp16")];
            int32 var_465 = const()[name = string("op_465"), val = int32(-2)];
            bool var_466_interleave_0 = const()[name = string("op_466_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_466_cast_fp16 = concat(axis = var_465, interleave = var_466_interleave_0, values = (var_463_cast_fp16, var_461_cast_fp16_0))[name = string("op_466_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_467_cast_fp16 = mul(x = var_466_cast_fp16, y = var_315_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_5_cast_fp16 = add(x = var_460_cast_fp16, y = var_467_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_54_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_54 = read_state(input = key_cache)[name = string("coreml_update_state_54")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_443_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_55_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_55 = read_state(input = value_cache)[name = string("coreml_update_state_55")];
            tensor<int32, [4]> var_537_begin_0 = const()[name = string("op_537_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_537_end_0 = const()[name = string("op_537_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_537_end_mask_0 = const()[name = string("op_537_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_537_cast_fp16 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = coreml_update_state_54)[name = string("op_537_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_540_axis_0 = const()[name = string("op_540_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_1 = split(axis = var_540_axis_0, split_sizes = tile_0, x = var_537_cast_fp16)[name = string("op_540_cast_fp16")];
            tensor<int32, [4]> var_547_begin_0 = const()[name = string("op_547_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_547_end_0 = const()[name = string("op_547_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_547_end_mask_0 = const()[name = string("op_547_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_547_cast_fp16 = slice_by_index(begin = var_547_begin_0, end = var_547_end_0, end_mask = var_547_end_mask_0, x = coreml_update_state_55)[name = string("op_547_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_550_axis_0 = const()[name = string("op_550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_1 = split(axis = var_550_axis_0, split_sizes = tile_1, x = var_547_cast_fp16)[name = string("op_550_cast_fp16")];
            tensor<int32, [2]> var_553_split_sizes_0 = const()[name = string("op_553_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_553_axis_0 = const()[name = string("op_553_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_553_0, tensor<fp16, [1, 8, 128, 16]> var_553_1 = split(axis = var_553_axis_0, split_sizes = var_553_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_553")];
            bool attn_weights_1_transpose_x_0 = const()[name = string("attn_weights_1_transpose_x_0"), val = bool(false)];
            bool attn_weights_1_transpose_y_0 = const()[name = string("attn_weights_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_540_cast_fp16_0, y = var_553_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_556_to_fp16 = const()[name = string("op_556_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_556_to_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_5_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
            int32 var_560 = const()[name = string("op_560"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_7_cast_fp16 = softmax(axis = var_560, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_566_transpose_x_1 = const()[name = string("op_566_transpose_x_1"), val = bool(true)];
            bool var_566_transpose_y_1 = const()[name = string("op_566_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_566_cast_fp16 = matmul(transpose_x = var_566_transpose_x_1, transpose_y = var_566_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_550_cast_fp16_0)[name = string("op_566_cast_fp16")];
            bool attn_weights_9_transpose_x_0 = const()[name = string("attn_weights_9_transpose_x_0"), val = bool(false)];
            bool attn_weights_9_transpose_y_0 = const()[name = string("attn_weights_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_9_cast_fp16 = matmul(transpose_x = attn_weights_9_transpose_x_0, transpose_y = attn_weights_9_transpose_y_0, x = var_540_cast_fp16_1, y = var_553_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_568_to_fp16 = const()[name = string("op_568_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_11_cast_fp16 = mul(x = attn_weights_9_cast_fp16, y = var_568_to_fp16)[name = string("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_13_cast_fp16 = add(x = attn_weights_11_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_13_cast_fp16")];
            int32 var_572 = const()[name = string("op_572"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_15_cast_fp16 = softmax(axis = var_572, x = attn_weights_13_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(true)];
            bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_15_cast_fp16, y = var_550_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_580 = const()[name = string("op_580"), val = int32(1)];
            bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_3_cast_fp16 = concat(axis = var_580, interleave = attn_output_3_interleave_0, values = (var_566_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_584_perm_0 = const()[name = string("op_584_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_584_cast_fp16 = transpose(perm = var_584_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_584_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332810496)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_617_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_617_cast_fp16")];
            int32 var_615 = const()[name = string("op_615"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_5_cast_fp16 = concat(axis = var_615, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_617_cast_fp16))[name = string("doubled_5_cast_fp16")];
            tensor<int32, [1]> out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_3_gamma_0_to_fp16 = const()[name = string("out_3_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341199168)))];
            fp16 var_627_to_fp16 = const()[name = string("op_627_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_627_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_638_split_sizes_0 = const()[name = string("op_638_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_638_axis_0 = const()[name = string("op_638_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_638_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_638_cast_fp16_1 = split(axis = var_638_axis_0, split_sizes = var_638_split_sizes_0, x = out_3_cast_fp16)[name = string("op_638_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_655_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_655_cast_fp16")];
            tensor<int32, [2]> var_661_strides_0 = const()[name = string("op_661_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_661_pad_type_0 = const()[name = string("op_661_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_661_pad_0 = const()[name = string("op_661_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_661_dilations_0 = const()[name = string("op_661_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_661_groups_0 = const()[name = string("op_661_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_661_cast_fp16 = conv(dilations = var_661_dilations_0, groups = var_661_groups_0, pad = var_661_pad_0, pad_type = var_661_pad_type_0, strides = var_661_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("op_661_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_9_cast_fp16 = mul(x = var_655_cast_fp16, y = var_661_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_679_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_679_cast_fp16")];
            int32 var_677 = const()[name = string("op_677"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_9_cast_fp16 = concat(axis = var_677, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_679_cast_fp16))[name = string("doubled_9_cast_fp16")];
            tensor<int32, [1]> out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_5_gamma_0_to_fp16 = const()[name = string("out_5_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341207424)))];
            fp16 var_689_to_fp16 = const()[name = string("op_689_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_689_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_700_split_sizes_0 = const()[name = string("op_700_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_700_axis_0 = const()[name = string("op_700_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_700_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_700_cast_fp16_1 = split(axis = var_700_axis_0, split_sizes = var_700_split_sizes_0, x = out_5_cast_fp16)[name = string("op_700_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341215680)))];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = var_700_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_757_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_757_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_764_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_764_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_768_cast_fp16 = mul(x = x_11_cast_fp16, y = var_308_cast_fp16)[name = string("op_768_cast_fp16")];
            tensor<int32, [2]> var_769_split_sizes_0 = const()[name = string("op_769_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_769_axis_0 = const()[name = string("op_769_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_769_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_769_cast_fp16_1 = split(axis = var_769_axis_0, split_sizes = var_769_split_sizes_0, x = x_11_cast_fp16)[name = string("op_769_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_771_cast_fp16 = mul(x = var_769_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_771_cast_fp16")];
            int32 var_773 = const()[name = string("op_773"), val = int32(-2)];
            bool var_774_interleave_0 = const()[name = string("op_774_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_774_cast_fp16 = concat(axis = var_773, interleave = var_774_interleave_0, values = (var_771_cast_fp16, var_769_cast_fp16_0))[name = string("op_774_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_775_cast_fp16 = mul(x = var_774_cast_fp16, y = var_315_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_9_cast_fp16 = add(x = var_768_cast_fp16, y = var_775_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_781_cast_fp16 = mul(x = var_757_cast_fp16, y = var_308_cast_fp16)[name = string("op_781_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_782_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = var_757_cast_fp16)[name = string("op_782_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_784_cast_fp16 = mul(x = var_782_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_784_cast_fp16")];
            int32 var_786 = const()[name = string("op_786"), val = int32(-2)];
            bool var_787_interleave_0 = const()[name = string("op_787_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_787_cast_fp16 = concat(axis = var_786, interleave = var_787_interleave_0, values = (var_784_cast_fp16, var_782_cast_fp16_0))[name = string("op_787_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_788_cast_fp16 = mul(x = var_787_cast_fp16, y = var_315_cast_fp16)[name = string("op_788_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_15_cast_fp16 = add(x = var_781_cast_fp16, y = var_788_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_54)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_56_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_56 = read_state(input = key_cache)[name = string("coreml_update_state_56")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_764_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_55)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_57_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_57 = read_state(input = value_cache)[name = string("coreml_update_state_57")];
            tensor<int32, [4]> var_858_begin_0 = const()[name = string("op_858_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_858_end_0 = const()[name = string("op_858_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_858_end_mask_0 = const()[name = string("op_858_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_858_cast_fp16 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = coreml_update_state_56)[name = string("op_858_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_861_axis_0 = const()[name = string("op_861_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_1 = split(axis = var_861_axis_0, split_sizes = tile_2, x = var_858_cast_fp16)[name = string("op_861_cast_fp16")];
            tensor<int32, [4]> var_868_begin_0 = const()[name = string("op_868_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_868_end_0 = const()[name = string("op_868_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_868_end_mask_0 = const()[name = string("op_868_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_868_cast_fp16 = slice_by_index(begin = var_868_begin_0, end = var_868_end_0, end_mask = var_868_end_mask_0, x = coreml_update_state_57)[name = string("op_868_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_871_axis_0 = const()[name = string("op_871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_1 = split(axis = var_871_axis_0, split_sizes = tile_3, x = var_868_cast_fp16)[name = string("op_871_cast_fp16")];
            tensor<int32, [2]> var_874_split_sizes_0 = const()[name = string("op_874_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_874_axis_0 = const()[name = string("op_874_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_874_0, tensor<fp16, [1, 8, 128, 16]> var_874_1 = split(axis = var_874_axis_0, split_sizes = var_874_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_874")];
            bool attn_weights_17_transpose_x_0 = const()[name = string("attn_weights_17_transpose_x_0"), val = bool(false)];
            bool attn_weights_17_transpose_y_0 = const()[name = string("attn_weights_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_17_cast_fp16 = matmul(transpose_x = attn_weights_17_transpose_x_0, transpose_y = attn_weights_17_transpose_y_0, x = var_861_cast_fp16_0, y = var_874_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_877_to_fp16 = const()[name = string("op_877_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_19_cast_fp16 = mul(x = attn_weights_17_cast_fp16, y = var_877_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            int32 var_881 = const()[name = string("op_881"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_23_cast_fp16 = softmax(axis = var_881, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_887_transpose_x_1 = const()[name = string("op_887_transpose_x_1"), val = bool(true)];
            bool var_887_transpose_y_1 = const()[name = string("op_887_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_887_cast_fp16 = matmul(transpose_x = var_887_transpose_x_1, transpose_y = var_887_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_871_cast_fp16_0)[name = string("op_887_cast_fp16")];
            bool attn_weights_25_transpose_x_0 = const()[name = string("attn_weights_25_transpose_x_0"), val = bool(false)];
            bool attn_weights_25_transpose_y_0 = const()[name = string("attn_weights_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_0, transpose_y = attn_weights_25_transpose_y_0, x = var_861_cast_fp16_1, y = var_874_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_889_to_fp16 = const()[name = string("op_889_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_889_to_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_29_cast_fp16 = add(x = attn_weights_27_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
            int32 var_893 = const()[name = string("op_893"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_31_cast_fp16 = softmax(axis = var_893, x = attn_weights_29_cast_fp16)[name = string("attn_weights_31_cast_fp16")];
            bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(true)];
            bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_31_cast_fp16, y = var_871_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_901 = const()[name = string("op_901"), val = int32(1)];
            bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_11_cast_fp16 = concat(axis = var_901, interleave = attn_output_11_interleave_0, values = (var_887_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_905_perm_0 = const()[name = string("op_905_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_905_cast_fp16 = transpose(perm = var_905_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_905_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_938_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_938_cast_fp16")];
            int32 var_936 = const()[name = string("op_936"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_13_cast_fp16 = concat(axis = var_936, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_938_cast_fp16))[name = string("doubled_13_cast_fp16")];
            tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_7_gamma_0_to_fp16 = const()[name = string("out_7_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342264320)))];
            fp16 var_948_to_fp16 = const()[name = string("op_948_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_948_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_959_split_sizes_0 = const()[name = string("op_959_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_959_axis_0 = const()[name = string("op_959_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_959_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_959_cast_fp16_1 = split(axis = var_959_axis_0, split_sizes = var_959_split_sizes_0, x = out_7_cast_fp16)[name = string("op_959_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_959_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_976_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_976_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342272576)))];
            tensor<int32, [2]> var_982_strides_0 = const()[name = string("op_982_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_982_pad_type_0 = const()[name = string("op_982_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_982_pad_0 = const()[name = string("op_982_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_982_dilations_0 = const()[name = string("op_982_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_982_groups_0 = const()[name = string("op_982_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_982_cast_fp16 = conv(dilations = var_982_dilations_0, groups = var_982_groups_0, pad = var_982_pad_0, pad_type = var_982_pad_type_0, strides = var_982_strides_0, weight = layers_1_mlp_up_proj_weight_to_fp16, x = var_959_cast_fp16_0)[name = string("op_982_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_19_cast_fp16 = mul(x = var_976_cast_fp16, y = var_982_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1000_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1000_cast_fp16")];
            int32 var_998 = const()[name = string("op_998"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_17_cast_fp16 = concat(axis = var_998, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1000_cast_fp16))[name = string("doubled_17_cast_fp16")];
            tensor<int32, [1]> out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_9_gamma_0_to_fp16 = const()[name = string("out_9_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367438464)))];
            fp16 var_1010_to_fp16 = const()[name = string("op_1010_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1010_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1021_split_sizes_0 = const()[name = string("op_1021_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1021_axis_0 = const()[name = string("op_1021_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1021_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1021_cast_fp16_1 = split(axis = var_1021_axis_0, split_sizes = var_1021_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1021_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367446720)))];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = var_1021_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_1078_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1078_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_1085_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1085_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1089_cast_fp16 = mul(x = x_21_cast_fp16, y = var_308_cast_fp16)[name = string("op_1089_cast_fp16")];
            tensor<int32, [2]> var_1090_split_sizes_0 = const()[name = string("op_1090_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1090_axis_0 = const()[name = string("op_1090_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_1090_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_1090_cast_fp16_1 = split(axis = var_1090_axis_0, split_sizes = var_1090_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1090_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_1092_cast_fp16 = mul(x = var_1090_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            int32 var_1094 = const()[name = string("op_1094"), val = int32(-2)];
            bool var_1095_interleave_0 = const()[name = string("op_1095_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_1095_cast_fp16 = concat(axis = var_1094, interleave = var_1095_interleave_0, values = (var_1092_cast_fp16, var_1090_cast_fp16_0))[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1096_cast_fp16 = mul(x = var_1095_cast_fp16, y = var_315_cast_fp16)[name = string("op_1096_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_15_cast_fp16 = add(x = var_1089_cast_fp16, y = var_1096_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1102_cast_fp16 = mul(x = var_1078_cast_fp16, y = var_308_cast_fp16)[name = string("op_1102_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_1103_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = var_1078_cast_fp16)[name = string("op_1103_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_1105_cast_fp16 = mul(x = var_1103_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1105_cast_fp16")];
            int32 var_1107 = const()[name = string("op_1107"), val = int32(-2)];
            bool var_1108_interleave_0 = const()[name = string("op_1108_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_1108_cast_fp16 = concat(axis = var_1107, interleave = var_1108_interleave_0, values = (var_1105_cast_fp16, var_1103_cast_fp16_0))[name = string("op_1108_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1109_cast_fp16 = mul(x = var_1108_cast_fp16, y = var_315_cast_fp16)[name = string("op_1109_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_25_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1109_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_56)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_58_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_58 = read_state(input = key_cache)[name = string("coreml_update_state_58")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1085_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_57)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_59_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_59 = read_state(input = value_cache)[name = string("coreml_update_state_59")];
            tensor<int32, [4]> var_1179_begin_0 = const()[name = string("op_1179_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1179_end_0 = const()[name = string("op_1179_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1179_end_mask_0 = const()[name = string("op_1179_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1179_cast_fp16 = slice_by_index(begin = var_1179_begin_0, end = var_1179_end_0, end_mask = var_1179_end_mask_0, x = coreml_update_state_58)[name = string("op_1179_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1182_axis_0 = const()[name = string("op_1182_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_1 = split(axis = var_1182_axis_0, split_sizes = tile_4, x = var_1179_cast_fp16)[name = string("op_1182_cast_fp16")];
            tensor<int32, [4]> var_1189_begin_0 = const()[name = string("op_1189_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1189_end_0 = const()[name = string("op_1189_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1189_end_mask_0 = const()[name = string("op_1189_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1189_cast_fp16 = slice_by_index(begin = var_1189_begin_0, end = var_1189_end_0, end_mask = var_1189_end_mask_0, x = coreml_update_state_59)[name = string("op_1189_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1192_axis_0 = const()[name = string("op_1192_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_1 = split(axis = var_1192_axis_0, split_sizes = tile_5, x = var_1189_cast_fp16)[name = string("op_1192_cast_fp16")];
            tensor<int32, [2]> var_1195_split_sizes_0 = const()[name = string("op_1195_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1195_axis_0 = const()[name = string("op_1195_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1195_0, tensor<fp16, [1, 8, 128, 16]> var_1195_1 = split(axis = var_1195_axis_0, split_sizes = var_1195_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1195")];
            bool attn_weights_33_transpose_x_0 = const()[name = string("attn_weights_33_transpose_x_0"), val = bool(false)];
            bool attn_weights_33_transpose_y_0 = const()[name = string("attn_weights_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_33_cast_fp16 = matmul(transpose_x = attn_weights_33_transpose_x_0, transpose_y = attn_weights_33_transpose_y_0, x = var_1182_cast_fp16_0, y = var_1195_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1198_to_fp16 = const()[name = string("op_1198_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_35_cast_fp16 = mul(x = attn_weights_33_cast_fp16, y = var_1198_to_fp16)[name = string("attn_weights_35_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_37_cast_fp16 = add(x = attn_weights_35_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_37_cast_fp16")];
            int32 var_1202 = const()[name = string("op_1202"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_39_cast_fp16 = softmax(axis = var_1202, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1208_transpose_x_1 = const()[name = string("op_1208_transpose_x_1"), val = bool(true)];
            bool var_1208_transpose_y_1 = const()[name = string("op_1208_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1208_cast_fp16 = matmul(transpose_x = var_1208_transpose_x_1, transpose_y = var_1208_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1192_cast_fp16_0)[name = string("op_1208_cast_fp16")];
            bool attn_weights_41_transpose_x_0 = const()[name = string("attn_weights_41_transpose_x_0"), val = bool(false)];
            bool attn_weights_41_transpose_y_0 = const()[name = string("attn_weights_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_41_cast_fp16 = matmul(transpose_x = attn_weights_41_transpose_x_0, transpose_y = attn_weights_41_transpose_y_0, x = var_1182_cast_fp16_1, y = var_1195_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1210_to_fp16 = const()[name = string("op_1210_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_43_cast_fp16 = mul(x = attn_weights_41_cast_fp16, y = var_1210_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            int32 var_1214 = const()[name = string("op_1214"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_47_cast_fp16 = softmax(axis = var_1214, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")];
            bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(true)];
            bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_47_cast_fp16, y = var_1192_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1222 = const()[name = string("op_1222"), val = int32(1)];
            bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_19_cast_fp16 = concat(axis = var_1222, interleave = attn_output_19_interleave_0, values = (var_1208_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1226_perm_0 = const()[name = string("op_1226_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_1226_cast_fp16 = transpose(perm = var_1226_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1226_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1259_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            int32 var_1257 = const()[name = string("op_1257"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_21_cast_fp16 = concat(axis = var_1257, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1259_cast_fp16))[name = string("doubled_21_cast_fp16")];
            tensor<int32, [1]> out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_11_gamma_0_to_fp16 = const()[name = string("out_11_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368495360)))];
            fp16 var_1269_to_fp16 = const()[name = string("op_1269_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1269_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1280_split_sizes_0 = const()[name = string("op_1280_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1280_axis_0 = const()[name = string("op_1280_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1280_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1280_cast_fp16_1 = split(axis = var_1280_axis_0, split_sizes = var_1280_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1280_cast_fp16")];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_1297_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1297_cast_fp16")];
            tensor<int32, [2]> var_1303_strides_0 = const()[name = string("op_1303_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1303_pad_type_0 = const()[name = string("op_1303_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1303_pad_0 = const()[name = string("op_1303_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1303_dilations_0 = const()[name = string("op_1303_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1303_groups_0 = const()[name = string("op_1303_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_1303_cast_fp16 = conv(dilations = var_1303_dilations_0, groups = var_1303_groups_0, pad = var_1303_pad_0, pad_type = var_1303_pad_type_0, strides = var_1303_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("op_1303_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_29_cast_fp16 = mul(x = var_1297_cast_fp16, y = var_1303_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1321_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1321_cast_fp16")];
            int32 var_1319 = const()[name = string("op_1319"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_25_cast_fp16 = concat(axis = var_1319, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1321_cast_fp16))[name = string("doubled_25_cast_fp16")];
            tensor<int32, [1]> out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_13_gamma_0_to_fp16 = const()[name = string("out_13_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368503616)))];
            fp16 var_1331_to_fp16 = const()[name = string("op_1331_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1331_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1342_split_sizes_0 = const()[name = string("op_1342_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1342_axis_0 = const()[name = string("op_1342_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1342_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1342_cast_fp16_1 = split(axis = var_1342_axis_0, split_sizes = var_1342_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1342_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368511872)))];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = var_1342_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_1399_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_1406_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1406_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1410_cast_fp16 = mul(x = x_31_cast_fp16, y = var_308_cast_fp16)[name = string("op_1410_cast_fp16")];
            tensor<int32, [2]> var_1411_split_sizes_0 = const()[name = string("op_1411_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1411_axis_0 = const()[name = string("op_1411_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_1411_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_1411_cast_fp16_1 = split(axis = var_1411_axis_0, split_sizes = var_1411_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1411_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_1413_cast_fp16 = mul(x = var_1411_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1413_cast_fp16")];
            int32 var_1415 = const()[name = string("op_1415"), val = int32(-2)];
            bool var_1416_interleave_0 = const()[name = string("op_1416_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_1416_cast_fp16 = concat(axis = var_1415, interleave = var_1416_interleave_0, values = (var_1413_cast_fp16, var_1411_cast_fp16_0))[name = string("op_1416_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1417_cast_fp16 = mul(x = var_1416_cast_fp16, y = var_315_cast_fp16)[name = string("op_1417_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_21_cast_fp16 = add(x = var_1410_cast_fp16, y = var_1417_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1423_cast_fp16 = mul(x = var_1399_cast_fp16, y = var_308_cast_fp16)[name = string("op_1423_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_1424_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = var_1399_cast_fp16)[name = string("op_1424_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_1426_cast_fp16 = mul(x = var_1424_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1426_cast_fp16")];
            int32 var_1428 = const()[name = string("op_1428"), val = int32(-2)];
            bool var_1429_interleave_0 = const()[name = string("op_1429_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_1429_cast_fp16 = concat(axis = var_1428, interleave = var_1429_interleave_0, values = (var_1426_cast_fp16, var_1424_cast_fp16_0))[name = string("op_1429_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1430_cast_fp16 = mul(x = var_1429_cast_fp16, y = var_315_cast_fp16)[name = string("op_1430_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_35_cast_fp16 = add(x = var_1423_cast_fp16, y = var_1430_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_58)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_60_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_60 = read_state(input = key_cache)[name = string("coreml_update_state_60")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1406_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_59)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_61_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_61 = read_state(input = value_cache)[name = string("coreml_update_state_61")];
            tensor<int32, [4]> var_1500_begin_0 = const()[name = string("op_1500_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1500_end_0 = const()[name = string("op_1500_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1500_end_mask_0 = const()[name = string("op_1500_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1500_cast_fp16 = slice_by_index(begin = var_1500_begin_0, end = var_1500_end_0, end_mask = var_1500_end_mask_0, x = coreml_update_state_60)[name = string("op_1500_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1503_axis_0 = const()[name = string("op_1503_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_1 = split(axis = var_1503_axis_0, split_sizes = tile_6, x = var_1500_cast_fp16)[name = string("op_1503_cast_fp16")];
            tensor<int32, [4]> var_1510_begin_0 = const()[name = string("op_1510_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1510_end_0 = const()[name = string("op_1510_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1510_end_mask_0 = const()[name = string("op_1510_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1510_cast_fp16 = slice_by_index(begin = var_1510_begin_0, end = var_1510_end_0, end_mask = var_1510_end_mask_0, x = coreml_update_state_61)[name = string("op_1510_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1513_axis_0 = const()[name = string("op_1513_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_1 = split(axis = var_1513_axis_0, split_sizes = tile_7, x = var_1510_cast_fp16)[name = string("op_1513_cast_fp16")];
            tensor<int32, [2]> var_1516_split_sizes_0 = const()[name = string("op_1516_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1516_axis_0 = const()[name = string("op_1516_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1516_0, tensor<fp16, [1, 8, 128, 16]> var_1516_1 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1516")];
            bool attn_weights_49_transpose_x_0 = const()[name = string("attn_weights_49_transpose_x_0"), val = bool(false)];
            bool attn_weights_49_transpose_y_0 = const()[name = string("attn_weights_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_0, transpose_y = attn_weights_49_transpose_y_0, x = var_1503_cast_fp16_0, y = var_1516_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_1519_to_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_53_cast_fp16 = add(x = attn_weights_51_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_53_cast_fp16")];
            int32 var_1523 = const()[name = string("op_1523"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_55_cast_fp16 = softmax(axis = var_1523, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1529_transpose_x_1 = const()[name = string("op_1529_transpose_x_1"), val = bool(true)];
            bool var_1529_transpose_y_1 = const()[name = string("op_1529_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1529_cast_fp16 = matmul(transpose_x = var_1529_transpose_x_1, transpose_y = var_1529_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1513_cast_fp16_0)[name = string("op_1529_cast_fp16")];
            bool attn_weights_57_transpose_x_0 = const()[name = string("attn_weights_57_transpose_x_0"), val = bool(false)];
            bool attn_weights_57_transpose_y_0 = const()[name = string("attn_weights_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_57_cast_fp16 = matmul(transpose_x = attn_weights_57_transpose_x_0, transpose_y = attn_weights_57_transpose_y_0, x = var_1503_cast_fp16_1, y = var_1516_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1531_to_fp16 = const()[name = string("op_1531_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_59_cast_fp16 = mul(x = attn_weights_57_cast_fp16, y = var_1531_to_fp16)[name = string("attn_weights_59_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_61_cast_fp16 = add(x = attn_weights_59_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_61_cast_fp16")];
            int32 var_1535 = const()[name = string("op_1535"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_63_cast_fp16 = softmax(axis = var_1535, x = attn_weights_61_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(true)];
            bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_63_cast_fp16, y = var_1513_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1543 = const()[name = string("op_1543"), val = int32(1)];
            bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_27_cast_fp16 = concat(axis = var_1543, interleave = attn_output_27_interleave_0, values = (var_1529_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1547_perm_0 = const()[name = string("op_1547_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_1547_cast_fp16 = transpose(perm = var_1547_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1547_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369560512)))];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1580_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1580_cast_fp16")];
            int32 var_1578 = const()[name = string("op_1578"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_29_cast_fp16 = concat(axis = var_1578, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1580_cast_fp16))[name = string("doubled_29_cast_fp16")];
            tensor<int32, [1]> out_15_axes_0 = const()[name = string("out_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_15_gamma_0_to_fp16 = const()[name = string("out_15_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377949184)))];
            fp16 var_1590_to_fp16 = const()[name = string("op_1590_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_1590_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1601_split_sizes_0 = const()[name = string("op_1601_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1601_axis_0 = const()[name = string("op_1601_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1601_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1601_cast_fp16_1 = split(axis = var_1601_axis_0, split_sizes = var_1601_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1601_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_1618_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1618_cast_fp16")];
            tensor<int32, [2]> var_1624_strides_0 = const()[name = string("op_1624_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1624_pad_type_0 = const()[name = string("op_1624_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1624_pad_0 = const()[name = string("op_1624_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1624_dilations_0 = const()[name = string("op_1624_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1624_groups_0 = const()[name = string("op_1624_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_1624_cast_fp16 = conv(dilations = var_1624_dilations_0, groups = var_1624_groups_0, pad = var_1624_pad_0, pad_type = var_1624_pad_type_0, strides = var_1624_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("op_1624_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_39_cast_fp16 = mul(x = var_1618_cast_fp16, y = var_1624_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1642_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1642_cast_fp16")];
            int32 var_1640 = const()[name = string("op_1640"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_33_cast_fp16 = concat(axis = var_1640, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1642_cast_fp16))[name = string("doubled_33_cast_fp16")];
            tensor<int32, [1]> out_17_axes_0 = const()[name = string("out_17_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_17_gamma_0_to_fp16 = const()[name = string("out_17_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377957440)))];
            fp16 var_1652_to_fp16 = const()[name = string("op_1652_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1652_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1663_split_sizes_0 = const()[name = string("op_1663_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1663_axis_0 = const()[name = string("op_1663_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1663_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1663_cast_fp16_1 = split(axis = var_1663_axis_0, split_sizes = var_1663_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1663_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377965696)))];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = var_1663_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_1720_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1720_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_1727_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1731_cast_fp16 = mul(x = x_41_cast_fp16, y = var_308_cast_fp16)[name = string("op_1731_cast_fp16")];
            tensor<int32, [2]> var_1732_split_sizes_0 = const()[name = string("op_1732_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1732_axis_0 = const()[name = string("op_1732_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_1732_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_1732_cast_fp16_1 = split(axis = var_1732_axis_0, split_sizes = var_1732_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1732_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_1734_cast_fp16 = mul(x = var_1732_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1734_cast_fp16")];
            int32 var_1736 = const()[name = string("op_1736"), val = int32(-2)];
            bool var_1737_interleave_0 = const()[name = string("op_1737_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_1737_cast_fp16 = concat(axis = var_1736, interleave = var_1737_interleave_0, values = (var_1734_cast_fp16, var_1732_cast_fp16_0))[name = string("op_1737_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1738_cast_fp16 = mul(x = var_1737_cast_fp16, y = var_315_cast_fp16)[name = string("op_1738_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_27_cast_fp16 = add(x = var_1731_cast_fp16, y = var_1738_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1744_cast_fp16 = mul(x = var_1720_cast_fp16, y = var_308_cast_fp16)[name = string("op_1744_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_1745_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = var_1720_cast_fp16)[name = string("op_1745_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_1747_cast_fp16 = mul(x = var_1745_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1747_cast_fp16")];
            int32 var_1749 = const()[name = string("op_1749"), val = int32(-2)];
            bool var_1750_interleave_0 = const()[name = string("op_1750_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_1750_cast_fp16 = concat(axis = var_1749, interleave = var_1750_interleave_0, values = (var_1747_cast_fp16, var_1745_cast_fp16_0))[name = string("op_1750_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1751_cast_fp16 = mul(x = var_1750_cast_fp16, y = var_315_cast_fp16)[name = string("op_1751_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_45_cast_fp16 = add(x = var_1744_cast_fp16, y = var_1751_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_60)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_62_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_62 = read_state(input = key_cache)[name = string("coreml_update_state_62")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1727_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_61)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_63_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_63 = read_state(input = value_cache)[name = string("coreml_update_state_63")];
            tensor<int32, [4]> var_1821_begin_0 = const()[name = string("op_1821_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1821_end_0 = const()[name = string("op_1821_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1821_end_mask_0 = const()[name = string("op_1821_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1821_cast_fp16 = slice_by_index(begin = var_1821_begin_0, end = var_1821_end_0, end_mask = var_1821_end_mask_0, x = coreml_update_state_62)[name = string("op_1821_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1824_axis_0 = const()[name = string("op_1824_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_1 = split(axis = var_1824_axis_0, split_sizes = tile_8, x = var_1821_cast_fp16)[name = string("op_1824_cast_fp16")];
            tensor<int32, [4]> var_1831_begin_0 = const()[name = string("op_1831_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1831_end_0 = const()[name = string("op_1831_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1831_end_mask_0 = const()[name = string("op_1831_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1831_cast_fp16 = slice_by_index(begin = var_1831_begin_0, end = var_1831_end_0, end_mask = var_1831_end_mask_0, x = coreml_update_state_63)[name = string("op_1831_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1834_axis_0 = const()[name = string("op_1834_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_1 = split(axis = var_1834_axis_0, split_sizes = tile_9, x = var_1831_cast_fp16)[name = string("op_1834_cast_fp16")];
            tensor<int32, [2]> var_1837_split_sizes_0 = const()[name = string("op_1837_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1837_axis_0 = const()[name = string("op_1837_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1837_0, tensor<fp16, [1, 8, 128, 16]> var_1837_1 = split(axis = var_1837_axis_0, split_sizes = var_1837_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1837")];
            bool attn_weights_65_transpose_x_0 = const()[name = string("attn_weights_65_transpose_x_0"), val = bool(false)];
            bool attn_weights_65_transpose_y_0 = const()[name = string("attn_weights_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_65_cast_fp16 = matmul(transpose_x = attn_weights_65_transpose_x_0, transpose_y = attn_weights_65_transpose_y_0, x = var_1824_cast_fp16_0, y = var_1837_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1840_to_fp16 = const()[name = string("op_1840_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_67_cast_fp16 = mul(x = attn_weights_65_cast_fp16, y = var_1840_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            int32 var_1844 = const()[name = string("op_1844"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_71_cast_fp16 = softmax(axis = var_1844, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1850_transpose_x_1 = const()[name = string("op_1850_transpose_x_1"), val = bool(true)];
            bool var_1850_transpose_y_1 = const()[name = string("op_1850_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1850_cast_fp16 = matmul(transpose_x = var_1850_transpose_x_1, transpose_y = var_1850_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1834_cast_fp16_0)[name = string("op_1850_cast_fp16")];
            bool attn_weights_73_transpose_x_0 = const()[name = string("attn_weights_73_transpose_x_0"), val = bool(false)];
            bool attn_weights_73_transpose_y_0 = const()[name = string("attn_weights_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_0, transpose_y = attn_weights_73_transpose_y_0, x = var_1824_cast_fp16_1, y = var_1837_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1852_to_fp16 = const()[name = string("op_1852_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1852_to_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_77_cast_fp16 = add(x = attn_weights_75_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_77_cast_fp16")];
            int32 var_1856 = const()[name = string("op_1856"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_79_cast_fp16 = softmax(axis = var_1856, x = attn_weights_77_cast_fp16)[name = string("attn_weights_79_cast_fp16")];
            bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(true)];
            bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_79_cast_fp16, y = var_1834_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1864 = const()[name = string("op_1864"), val = int32(1)];
            bool attn_output_35_interleave_0 = const()[name = string("attn_output_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_35_cast_fp16 = concat(axis = var_1864, interleave = attn_output_35_interleave_0, values = (var_1850_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1868_perm_0 = const()[name = string("op_1868_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_1868_cast_fp16 = transpose(perm = var_1868_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1868_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1901_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1901_cast_fp16")];
            int32 var_1899 = const()[name = string("op_1899"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_37_cast_fp16 = concat(axis = var_1899, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1901_cast_fp16))[name = string("doubled_37_cast_fp16")];
            tensor<int32, [1]> out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_19_gamma_0_to_fp16 = const()[name = string("out_19_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379014336)))];
            fp16 var_1911_to_fp16 = const()[name = string("op_1911_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1911_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1922_split_sizes_0 = const()[name = string("op_1922_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1922_axis_0 = const()[name = string("op_1922_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1922_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1922_cast_fp16_1 = split(axis = var_1922_axis_0, split_sizes = var_1922_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1922_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_1939_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<int32, [2]> var_1945_strides_0 = const()[name = string("op_1945_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1945_pad_type_0 = const()[name = string("op_1945_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1945_pad_0 = const()[name = string("op_1945_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1945_dilations_0 = const()[name = string("op_1945_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1945_groups_0 = const()[name = string("op_1945_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_1945_cast_fp16 = conv(dilations = var_1945_dilations_0, groups = var_1945_groups_0, pad = var_1945_pad_0, pad_type = var_1945_pad_type_0, strides = var_1945_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("op_1945_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_49_cast_fp16 = mul(x = var_1939_cast_fp16, y = var_1945_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_1963_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1963_cast_fp16")];
            int32 var_1961 = const()[name = string("op_1961"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_41_cast_fp16 = concat(axis = var_1961, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1963_cast_fp16))[name = string("doubled_41_cast_fp16")];
            tensor<int32, [1]> out_21_axes_0 = const()[name = string("out_21_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_21_gamma_0_to_fp16 = const()[name = string("out_21_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379022592)))];
            fp16 var_1973_to_fp16 = const()[name = string("op_1973_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1973_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_1984_split_sizes_0 = const()[name = string("op_1984_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1984_axis_0 = const()[name = string("op_1984_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1984_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1984_cast_fp16_1 = split(axis = var_1984_axis_0, split_sizes = var_1984_split_sizes_0, x = out_21_cast_fp16)[name = string("op_1984_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379030848)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_1984_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_2041_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2041_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_2048_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2048_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2052_cast_fp16 = mul(x = x_51_cast_fp16, y = var_308_cast_fp16)[name = string("op_2052_cast_fp16")];
            tensor<int32, [2]> var_2053_split_sizes_0 = const()[name = string("op_2053_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2053_axis_0 = const()[name = string("op_2053_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_2053_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_2053_cast_fp16_1 = split(axis = var_2053_axis_0, split_sizes = var_2053_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2053_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_2055_cast_fp16 = mul(x = var_2053_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2055_cast_fp16")];
            int32 var_2057 = const()[name = string("op_2057"), val = int32(-2)];
            bool var_2058_interleave_0 = const()[name = string("op_2058_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_2058_cast_fp16 = concat(axis = var_2057, interleave = var_2058_interleave_0, values = (var_2055_cast_fp16, var_2053_cast_fp16_0))[name = string("op_2058_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2059_cast_fp16 = mul(x = var_2058_cast_fp16, y = var_315_cast_fp16)[name = string("op_2059_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_33_cast_fp16 = add(x = var_2052_cast_fp16, y = var_2059_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2065_cast_fp16 = mul(x = var_2041_cast_fp16, y = var_308_cast_fp16)[name = string("op_2065_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_2066_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = var_2041_cast_fp16)[name = string("op_2066_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_2068_cast_fp16 = mul(x = var_2066_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2068_cast_fp16")];
            int32 var_2070 = const()[name = string("op_2070"), val = int32(-2)];
            bool var_2071_interleave_0 = const()[name = string("op_2071_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_2071_cast_fp16 = concat(axis = var_2070, interleave = var_2071_interleave_0, values = (var_2068_cast_fp16, var_2066_cast_fp16_0))[name = string("op_2071_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2072_cast_fp16 = mul(x = var_2071_cast_fp16, y = var_315_cast_fp16)[name = string("op_2072_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_55_cast_fp16 = add(x = var_2065_cast_fp16, y = var_2072_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_62)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_64_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_64 = read_state(input = key_cache)[name = string("coreml_update_state_64")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2048_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_63)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_65_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_65 = read_state(input = value_cache)[name = string("coreml_update_state_65")];
            tensor<int32, [4]> var_2142_begin_0 = const()[name = string("op_2142_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2142_end_0 = const()[name = string("op_2142_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2142_end_mask_0 = const()[name = string("op_2142_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2142_cast_fp16 = slice_by_index(begin = var_2142_begin_0, end = var_2142_end_0, end_mask = var_2142_end_mask_0, x = coreml_update_state_64)[name = string("op_2142_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2145_axis_0 = const()[name = string("op_2145_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_1 = split(axis = var_2145_axis_0, split_sizes = tile_10, x = var_2142_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [4]> var_2152_begin_0 = const()[name = string("op_2152_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2152_end_0 = const()[name = string("op_2152_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2152_end_mask_0 = const()[name = string("op_2152_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2152_cast_fp16 = slice_by_index(begin = var_2152_begin_0, end = var_2152_end_0, end_mask = var_2152_end_mask_0, x = coreml_update_state_65)[name = string("op_2152_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2155_axis_0 = const()[name = string("op_2155_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_1 = split(axis = var_2155_axis_0, split_sizes = tile_11, x = var_2152_cast_fp16)[name = string("op_2155_cast_fp16")];
            tensor<int32, [2]> var_2158_split_sizes_0 = const()[name = string("op_2158_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2158_axis_0 = const()[name = string("op_2158_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_2158_0, tensor<fp16, [1, 8, 128, 16]> var_2158_1 = split(axis = var_2158_axis_0, split_sizes = var_2158_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2158")];
            bool attn_weights_81_transpose_x_0 = const()[name = string("attn_weights_81_transpose_x_0"), val = bool(false)];
            bool attn_weights_81_transpose_y_0 = const()[name = string("attn_weights_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_81_cast_fp16 = matmul(transpose_x = attn_weights_81_transpose_x_0, transpose_y = attn_weights_81_transpose_y_0, x = var_2145_cast_fp16_0, y = var_2158_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2161_to_fp16 = const()[name = string("op_2161_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_83_cast_fp16 = mul(x = attn_weights_81_cast_fp16, y = var_2161_to_fp16)[name = string("attn_weights_83_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_85_cast_fp16 = add(x = attn_weights_83_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_85_cast_fp16")];
            int32 var_2165 = const()[name = string("op_2165"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_87_cast_fp16 = softmax(axis = var_2165, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2171_transpose_x_1 = const()[name = string("op_2171_transpose_x_1"), val = bool(true)];
            bool var_2171_transpose_y_1 = const()[name = string("op_2171_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_2171_cast_fp16 = matmul(transpose_x = var_2171_transpose_x_1, transpose_y = var_2171_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2155_cast_fp16_0)[name = string("op_2171_cast_fp16")];
            bool attn_weights_89_transpose_x_0 = const()[name = string("attn_weights_89_transpose_x_0"), val = bool(false)];
            bool attn_weights_89_transpose_y_0 = const()[name = string("attn_weights_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_89_cast_fp16 = matmul(transpose_x = attn_weights_89_transpose_x_0, transpose_y = attn_weights_89_transpose_y_0, x = var_2145_cast_fp16_1, y = var_2158_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2173_to_fp16 = const()[name = string("op_2173_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_91_cast_fp16 = mul(x = attn_weights_89_cast_fp16, y = var_2173_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            int32 var_2177 = const()[name = string("op_2177"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_95_cast_fp16 = softmax(axis = var_2177, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")];
            bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(true)];
            bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_95_cast_fp16, y = var_2155_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2185 = const()[name = string("op_2185"), val = int32(1)];
            bool attn_output_43_interleave_0 = const()[name = string("attn_output_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_43_cast_fp16 = concat(axis = var_2185, interleave = attn_output_43_interleave_0, values = (var_2171_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2189_perm_0 = const()[name = string("op_2189_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_2189_cast_fp16 = transpose(perm = var_2189_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2189_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_2222_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2222_cast_fp16")];
            int32 var_2220 = const()[name = string("op_2220"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_45_cast_fp16 = concat(axis = var_2220, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2222_cast_fp16))[name = string("doubled_45_cast_fp16")];
            tensor<int32, [1]> out_23_axes_0 = const()[name = string("out_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_23_gamma_0_to_fp16 = const()[name = string("out_23_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380079488)))];
            fp16 var_2232_to_fp16 = const()[name = string("op_2232_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_2232_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2243_split_sizes_0 = const()[name = string("op_2243_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2243_axis_0 = const()[name = string("op_2243_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2243_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2243_cast_fp16_1 = split(axis = var_2243_axis_0, split_sizes = var_2243_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2243_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_2260_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<int32, [2]> var_2266_strides_0 = const()[name = string("op_2266_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2266_pad_type_0 = const()[name = string("op_2266_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2266_pad_0 = const()[name = string("op_2266_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2266_dilations_0 = const()[name = string("op_2266_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2266_groups_0 = const()[name = string("op_2266_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_2266_cast_fp16 = conv(dilations = var_2266_dilations_0, groups = var_2266_groups_0, pad = var_2266_pad_0, pad_type = var_2266_pad_type_0, strides = var_2266_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("op_2266_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_59_cast_fp16 = mul(x = var_2260_cast_fp16, y = var_2266_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_2284_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2284_cast_fp16")];
            int32 var_2282 = const()[name = string("op_2282"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_49_cast_fp16 = concat(axis = var_2282, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2284_cast_fp16))[name = string("doubled_49_cast_fp16")];
            tensor<int32, [1]> out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_25_gamma_0_to_fp16 = const()[name = string("out_25_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380087744)))];
            fp16 var_2294_to_fp16 = const()[name = string("op_2294_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_2294_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2305_split_sizes_0 = const()[name = string("op_2305_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2305_axis_0 = const()[name = string("op_2305_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2305_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2305_cast_fp16_1 = split(axis = var_2305_axis_0, split_sizes = var_2305_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380096000)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2305_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_2362_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2362_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_2369_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2369_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2373_cast_fp16 = mul(x = x_61_cast_fp16, y = var_308_cast_fp16)[name = string("op_2373_cast_fp16")];
            tensor<int32, [2]> var_2374_split_sizes_0 = const()[name = string("op_2374_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2374_axis_0 = const()[name = string("op_2374_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_2374_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_2374_cast_fp16_1 = split(axis = var_2374_axis_0, split_sizes = var_2374_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2376_cast_fp16")];
            int32 var_2378 = const()[name = string("op_2378"), val = int32(-2)];
            bool var_2379_interleave_0 = const()[name = string("op_2379_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_2379_cast_fp16 = concat(axis = var_2378, interleave = var_2379_interleave_0, values = (var_2376_cast_fp16, var_2374_cast_fp16_0))[name = string("op_2379_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2380_cast_fp16 = mul(x = var_2379_cast_fp16, y = var_315_cast_fp16)[name = string("op_2380_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_39_cast_fp16 = add(x = var_2373_cast_fp16, y = var_2380_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2386_cast_fp16 = mul(x = var_2362_cast_fp16, y = var_308_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_2387_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = var_2362_cast_fp16)[name = string("op_2387_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_2389_cast_fp16 = mul(x = var_2387_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2389_cast_fp16")];
            int32 var_2391 = const()[name = string("op_2391"), val = int32(-2)];
            bool var_2392_interleave_0 = const()[name = string("op_2392_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_2392_cast_fp16 = concat(axis = var_2391, interleave = var_2392_interleave_0, values = (var_2389_cast_fp16, var_2387_cast_fp16_0))[name = string("op_2392_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2393_cast_fp16 = mul(x = var_2392_cast_fp16, y = var_315_cast_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_65_cast_fp16 = add(x = var_2386_cast_fp16, y = var_2393_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_64)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_66_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_66 = read_state(input = key_cache)[name = string("coreml_update_state_66")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2369_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_65)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_67_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_67 = read_state(input = value_cache)[name = string("coreml_update_state_67")];
            tensor<int32, [4]> var_2463_begin_0 = const()[name = string("op_2463_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2463_end_0 = const()[name = string("op_2463_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2463_end_mask_0 = const()[name = string("op_2463_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2463_cast_fp16 = slice_by_index(begin = var_2463_begin_0, end = var_2463_end_0, end_mask = var_2463_end_mask_0, x = coreml_update_state_66)[name = string("op_2463_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2466_axis_0 = const()[name = string("op_2466_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_1 = split(axis = var_2466_axis_0, split_sizes = tile_12, x = var_2463_cast_fp16)[name = string("op_2466_cast_fp16")];
            tensor<int32, [4]> var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, x = coreml_update_state_67)[name = string("op_2473_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2476_axis_0 = const()[name = string("op_2476_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_1 = split(axis = var_2476_axis_0, split_sizes = tile_13, x = var_2473_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<int32, [2]> var_2479_split_sizes_0 = const()[name = string("op_2479_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2479_axis_0 = const()[name = string("op_2479_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_2479_0, tensor<fp16, [1, 8, 128, 16]> var_2479_1 = split(axis = var_2479_axis_0, split_sizes = var_2479_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2479")];
            bool attn_weights_97_transpose_x_0 = const()[name = string("attn_weights_97_transpose_x_0"), val = bool(false)];
            bool attn_weights_97_transpose_y_0 = const()[name = string("attn_weights_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_0, transpose_y = attn_weights_97_transpose_y_0, x = var_2466_cast_fp16_0, y = var_2479_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2482_to_fp16 = const()[name = string("op_2482_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_2482_to_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_101_cast_fp16 = add(x = attn_weights_99_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_101_cast_fp16")];
            int32 var_2486 = const()[name = string("op_2486"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_103_cast_fp16 = softmax(axis = var_2486, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2492_transpose_x_1 = const()[name = string("op_2492_transpose_x_1"), val = bool(true)];
            bool var_2492_transpose_y_1 = const()[name = string("op_2492_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_2492_cast_fp16 = matmul(transpose_x = var_2492_transpose_x_1, transpose_y = var_2492_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2476_cast_fp16_0)[name = string("op_2492_cast_fp16")];
            bool attn_weights_105_transpose_x_0 = const()[name = string("attn_weights_105_transpose_x_0"), val = bool(false)];
            bool attn_weights_105_transpose_y_0 = const()[name = string("attn_weights_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_105_cast_fp16 = matmul(transpose_x = attn_weights_105_transpose_x_0, transpose_y = attn_weights_105_transpose_y_0, x = var_2466_cast_fp16_1, y = var_2479_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2494_to_fp16 = const()[name = string("op_2494_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_107_cast_fp16 = mul(x = attn_weights_105_cast_fp16, y = var_2494_to_fp16)[name = string("attn_weights_107_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_109_cast_fp16 = add(x = attn_weights_107_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_109_cast_fp16")];
            int32 var_2498 = const()[name = string("op_2498"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_111_cast_fp16 = softmax(axis = var_2498, x = attn_weights_109_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(true)];
            bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_111_cast_fp16, y = var_2476_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2506 = const()[name = string("op_2506"), val = int32(1)];
            bool attn_output_51_interleave_0 = const()[name = string("attn_output_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_51_cast_fp16 = concat(axis = var_2506, interleave = attn_output_51_interleave_0, values = (var_2492_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2510_perm_0 = const()[name = string("op_2510_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_2510_cast_fp16 = transpose(perm = var_2510_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2510_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_2543_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2543_cast_fp16")];
            int32 var_2541 = const()[name = string("op_2541"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_53_cast_fp16 = concat(axis = var_2541, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2543_cast_fp16))[name = string("doubled_53_cast_fp16")];
            tensor<int32, [1]> out_27_axes_0 = const()[name = string("out_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_27_gamma_0_to_fp16 = const()[name = string("out_27_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381144640)))];
            fp16 var_2553_to_fp16 = const()[name = string("op_2553_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_2553_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2564_split_sizes_0 = const()[name = string("op_2564_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2564_axis_0 = const()[name = string("op_2564_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2564_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2564_cast_fp16_1 = split(axis = var_2564_axis_0, split_sizes = var_2564_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_2581_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [2]> var_2587_strides_0 = const()[name = string("op_2587_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2587_pad_type_0 = const()[name = string("op_2587_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2587_pad_0 = const()[name = string("op_2587_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2587_dilations_0 = const()[name = string("op_2587_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2587_groups_0 = const()[name = string("op_2587_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_2587_cast_fp16 = conv(dilations = var_2587_dilations_0, groups = var_2587_groups_0, pad = var_2587_pad_0, pad_type = var_2587_pad_type_0, strides = var_2587_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("op_2587_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_69_cast_fp16 = mul(x = var_2581_cast_fp16, y = var_2587_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381152896)))];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_to_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_2605_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2605_cast_fp16")];
            int32 var_2603 = const()[name = string("op_2603"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_57_cast_fp16 = concat(axis = var_2603, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2605_cast_fp16))[name = string("doubled_57_cast_fp16")];
            tensor<int32, [1]> out_29_axes_0 = const()[name = string("out_29_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_29_gamma_0_to_fp16 = const()[name = string("out_29_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406318784)))];
            fp16 var_2615_to_fp16 = const()[name = string("op_2615_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_2615_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2626_split_sizes_0 = const()[name = string("op_2626_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2626_axis_0 = const()[name = string("op_2626_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2626_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2626_cast_fp16_1 = split(axis = var_2626_axis_0, split_sizes = var_2626_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2626_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406327040)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2626_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_2683_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_2690_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2690_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2694_cast_fp16 = mul(x = x_71_cast_fp16, y = var_308_cast_fp16)[name = string("op_2694_cast_fp16")];
            tensor<int32, [2]> var_2695_split_sizes_0 = const()[name = string("op_2695_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2695_axis_0 = const()[name = string("op_2695_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_2695_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_2695_cast_fp16_1 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2695_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_2697_cast_fp16 = mul(x = var_2695_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2697_cast_fp16")];
            int32 var_2699 = const()[name = string("op_2699"), val = int32(-2)];
            bool var_2700_interleave_0 = const()[name = string("op_2700_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_2700_cast_fp16 = concat(axis = var_2699, interleave = var_2700_interleave_0, values = (var_2697_cast_fp16, var_2695_cast_fp16_0))[name = string("op_2700_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2701_cast_fp16 = mul(x = var_2700_cast_fp16, y = var_315_cast_fp16)[name = string("op_2701_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_45_cast_fp16 = add(x = var_2694_cast_fp16, y = var_2701_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2707_cast_fp16 = mul(x = var_2683_cast_fp16, y = var_308_cast_fp16)[name = string("op_2707_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_2708_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = var_2683_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_2710_cast_fp16 = mul(x = var_2708_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2710_cast_fp16")];
            int32 var_2712 = const()[name = string("op_2712"), val = int32(-2)];
            bool var_2713_interleave_0 = const()[name = string("op_2713_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_2713_cast_fp16 = concat(axis = var_2712, interleave = var_2713_interleave_0, values = (var_2710_cast_fp16, var_2708_cast_fp16_0))[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2714_cast_fp16 = mul(x = var_2713_cast_fp16, y = var_315_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_75_cast_fp16 = add(x = var_2707_cast_fp16, y = var_2714_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_66)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_68_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_68 = read_state(input = key_cache)[name = string("coreml_update_state_68")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2690_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_67)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_69_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_69 = read_state(input = value_cache)[name = string("coreml_update_state_69")];
            tensor<int32, [4]> var_2784_begin_0 = const()[name = string("op_2784_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2784_end_0 = const()[name = string("op_2784_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2784_end_mask_0 = const()[name = string("op_2784_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2784_cast_fp16 = slice_by_index(begin = var_2784_begin_0, end = var_2784_end_0, end_mask = var_2784_end_mask_0, x = coreml_update_state_68)[name = string("op_2784_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2787_axis_0 = const()[name = string("op_2787_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_1 = split(axis = var_2787_axis_0, split_sizes = tile_14, x = var_2784_cast_fp16)[name = string("op_2787_cast_fp16")];
            tensor<int32, [4]> var_2794_begin_0 = const()[name = string("op_2794_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2794_end_0 = const()[name = string("op_2794_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2794_end_mask_0 = const()[name = string("op_2794_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2794_cast_fp16 = slice_by_index(begin = var_2794_begin_0, end = var_2794_end_0, end_mask = var_2794_end_mask_0, x = coreml_update_state_69)[name = string("op_2794_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2797_axis_0 = const()[name = string("op_2797_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_1 = split(axis = var_2797_axis_0, split_sizes = tile_15, x = var_2794_cast_fp16)[name = string("op_2797_cast_fp16")];
            tensor<int32, [2]> var_2800_split_sizes_0 = const()[name = string("op_2800_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2800_axis_0 = const()[name = string("op_2800_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_2800_0, tensor<fp16, [1, 8, 128, 16]> var_2800_1 = split(axis = var_2800_axis_0, split_sizes = var_2800_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2800")];
            bool attn_weights_113_transpose_x_0 = const()[name = string("attn_weights_113_transpose_x_0"), val = bool(false)];
            bool attn_weights_113_transpose_y_0 = const()[name = string("attn_weights_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_113_cast_fp16 = matmul(transpose_x = attn_weights_113_transpose_x_0, transpose_y = attn_weights_113_transpose_y_0, x = var_2787_cast_fp16_0, y = var_2800_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2803_to_fp16 = const()[name = string("op_2803_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_115_cast_fp16 = mul(x = attn_weights_113_cast_fp16, y = var_2803_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            int32 var_2807 = const()[name = string("op_2807"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_119_cast_fp16 = softmax(axis = var_2807, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2813_transpose_x_1 = const()[name = string("op_2813_transpose_x_1"), val = bool(true)];
            bool var_2813_transpose_y_1 = const()[name = string("op_2813_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_2813_cast_fp16 = matmul(transpose_x = var_2813_transpose_x_1, transpose_y = var_2813_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2797_cast_fp16_0)[name = string("op_2813_cast_fp16")];
            bool attn_weights_121_transpose_x_0 = const()[name = string("attn_weights_121_transpose_x_0"), val = bool(false)];
            bool attn_weights_121_transpose_y_0 = const()[name = string("attn_weights_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_0, transpose_y = attn_weights_121_transpose_y_0, x = var_2787_cast_fp16_1, y = var_2800_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2815_to_fp16 = const()[name = string("op_2815_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_2815_to_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_125_cast_fp16 = add(x = attn_weights_123_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_125_cast_fp16")];
            int32 var_2819 = const()[name = string("op_2819"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_127_cast_fp16 = softmax(axis = var_2819, x = attn_weights_125_cast_fp16)[name = string("attn_weights_127_cast_fp16")];
            bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(true)];
            bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_127_cast_fp16, y = var_2797_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2827 = const()[name = string("op_2827"), val = int32(1)];
            bool attn_output_59_interleave_0 = const()[name = string("attn_output_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_59_cast_fp16 = concat(axis = var_2827, interleave = attn_output_59_interleave_0, values = (var_2813_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2831_perm_0 = const()[name = string("op_2831_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_2831_cast_fp16 = transpose(perm = var_2831_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2831_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_2864_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2864_cast_fp16")];
            int32 var_2862 = const()[name = string("op_2862"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_61_cast_fp16 = concat(axis = var_2862, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2864_cast_fp16))[name = string("doubled_61_cast_fp16")];
            tensor<int32, [1]> out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_31_gamma_0_to_fp16 = const()[name = string("out_31_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407375680)))];
            fp16 var_2874_to_fp16 = const()[name = string("op_2874_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_2874_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2885_split_sizes_0 = const()[name = string("op_2885_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2885_axis_0 = const()[name = string("op_2885_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2885_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2885_cast_fp16_1 = split(axis = var_2885_axis_0, split_sizes = var_2885_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2885_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2885_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_2902_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2902_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407383936)))];
            tensor<int32, [2]> var_2908_strides_0 = const()[name = string("op_2908_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2908_pad_type_0 = const()[name = string("op_2908_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2908_pad_0 = const()[name = string("op_2908_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2908_dilations_0 = const()[name = string("op_2908_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2908_groups_0 = const()[name = string("op_2908_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_2908_cast_fp16 = conv(dilations = var_2908_dilations_0, groups = var_2908_groups_0, pad = var_2908_pad_0, pad_type = var_2908_pad_type_0, strides = var_2908_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2885_cast_fp16_0)[name = string("op_2908_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_79_cast_fp16 = mul(x = var_2902_cast_fp16, y = var_2908_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432549824)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_2926_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2926_cast_fp16")];
            int32 var_2924 = const()[name = string("op_2924"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_65_cast_fp16 = concat(axis = var_2924, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2926_cast_fp16))[name = string("doubled_65_cast_fp16")];
            tensor<int32, [1]> out_33_axes_0 = const()[name = string("out_33_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_33_gamma_0_to_fp16 = const()[name = string("out_33_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457715712)))];
            fp16 var_2936_to_fp16 = const()[name = string("op_2936_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_2936_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2947_split_sizes_0 = const()[name = string("op_2947_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2947_axis_0 = const()[name = string("op_2947_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2947_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2947_cast_fp16_1 = split(axis = var_2947_axis_0, split_sizes = var_2947_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2947_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457723968)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466112640)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467161280)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_3004_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3004_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_3011_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3011_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3015_cast_fp16 = mul(x = x_81_cast_fp16, y = var_308_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<int32, [2]> var_3016_split_sizes_0 = const()[name = string("op_3016_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3016_axis_0 = const()[name = string("op_3016_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_3016_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_3016_cast_fp16_1 = split(axis = var_3016_axis_0, split_sizes = var_3016_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3016_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_3018_cast_fp16 = mul(x = var_3016_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3018_cast_fp16")];
            int32 var_3020 = const()[name = string("op_3020"), val = int32(-2)];
            bool var_3021_interleave_0 = const()[name = string("op_3021_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_3021_cast_fp16 = concat(axis = var_3020, interleave = var_3021_interleave_0, values = (var_3018_cast_fp16, var_3016_cast_fp16_0))[name = string("op_3021_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3022_cast_fp16 = mul(x = var_3021_cast_fp16, y = var_315_cast_fp16)[name = string("op_3022_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_51_cast_fp16 = add(x = var_3015_cast_fp16, y = var_3022_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3028_cast_fp16 = mul(x = var_3004_cast_fp16, y = var_308_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_3029_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = var_3004_cast_fp16)[name = string("op_3029_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_3031_cast_fp16 = mul(x = var_3029_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3031_cast_fp16")];
            int32 var_3033 = const()[name = string("op_3033"), val = int32(-2)];
            bool var_3034_interleave_0 = const()[name = string("op_3034_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_3034_cast_fp16 = concat(axis = var_3033, interleave = var_3034_interleave_0, values = (var_3031_cast_fp16, var_3029_cast_fp16_0))[name = string("op_3034_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3035_cast_fp16 = mul(x = var_3034_cast_fp16, y = var_315_cast_fp16)[name = string("op_3035_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_85_cast_fp16 = add(x = var_3028_cast_fp16, y = var_3035_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_68)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_70_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_70 = read_state(input = key_cache)[name = string("coreml_update_state_70")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3011_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_69)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_71_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_71 = read_state(input = value_cache)[name = string("coreml_update_state_71")];
            tensor<int32, [4]> var_3105_begin_0 = const()[name = string("op_3105_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3105_end_0 = const()[name = string("op_3105_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3105_end_mask_0 = const()[name = string("op_3105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3105_cast_fp16 = slice_by_index(begin = var_3105_begin_0, end = var_3105_end_0, end_mask = var_3105_end_mask_0, x = coreml_update_state_70)[name = string("op_3105_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3108_axis_0 = const()[name = string("op_3108_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_1 = split(axis = var_3108_axis_0, split_sizes = tile_16, x = var_3105_cast_fp16)[name = string("op_3108_cast_fp16")];
            tensor<int32, [4]> var_3115_begin_0 = const()[name = string("op_3115_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3115_end_0 = const()[name = string("op_3115_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3115_end_mask_0 = const()[name = string("op_3115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3115_cast_fp16 = slice_by_index(begin = var_3115_begin_0, end = var_3115_end_0, end_mask = var_3115_end_mask_0, x = coreml_update_state_71)[name = string("op_3115_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3118_axis_0 = const()[name = string("op_3118_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_1 = split(axis = var_3118_axis_0, split_sizes = tile_17, x = var_3115_cast_fp16)[name = string("op_3118_cast_fp16")];
            tensor<int32, [2]> var_3121_split_sizes_0 = const()[name = string("op_3121_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3121_axis_0 = const()[name = string("op_3121_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_3121_0, tensor<fp16, [1, 8, 128, 16]> var_3121_1 = split(axis = var_3121_axis_0, split_sizes = var_3121_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3121")];
            bool attn_weights_129_transpose_x_0 = const()[name = string("attn_weights_129_transpose_x_0"), val = bool(false)];
            bool attn_weights_129_transpose_y_0 = const()[name = string("attn_weights_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_129_cast_fp16 = matmul(transpose_x = attn_weights_129_transpose_x_0, transpose_y = attn_weights_129_transpose_y_0, x = var_3108_cast_fp16_0, y = var_3121_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3124_to_fp16 = const()[name = string("op_3124_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_131_cast_fp16 = mul(x = attn_weights_129_cast_fp16, y = var_3124_to_fp16)[name = string("attn_weights_131_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_133_cast_fp16 = add(x = attn_weights_131_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_133_cast_fp16")];
            int32 var_3128 = const()[name = string("op_3128"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_135_cast_fp16 = softmax(axis = var_3128, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3134_transpose_x_1 = const()[name = string("op_3134_transpose_x_1"), val = bool(true)];
            bool var_3134_transpose_y_1 = const()[name = string("op_3134_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_3134_cast_fp16 = matmul(transpose_x = var_3134_transpose_x_1, transpose_y = var_3134_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3118_cast_fp16_0)[name = string("op_3134_cast_fp16")];
            bool attn_weights_137_transpose_x_0 = const()[name = string("attn_weights_137_transpose_x_0"), val = bool(false)];
            bool attn_weights_137_transpose_y_0 = const()[name = string("attn_weights_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_137_cast_fp16 = matmul(transpose_x = attn_weights_137_transpose_x_0, transpose_y = attn_weights_137_transpose_y_0, x = var_3108_cast_fp16_1, y = var_3121_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3136_to_fp16 = const()[name = string("op_3136_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_139_cast_fp16 = mul(x = attn_weights_137_cast_fp16, y = var_3136_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            int32 var_3140 = const()[name = string("op_3140"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_cast_fp16 = softmax(axis = var_3140, x = attn_weights_141_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(true)];
            bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_cast_fp16, y = var_3118_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3148 = const()[name = string("op_3148"), val = int32(1)];
            bool attn_output_67_interleave_0 = const()[name = string("attn_output_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_67_cast_fp16 = concat(axis = var_3148, interleave = attn_output_67_interleave_0, values = (var_3134_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3152_perm_0 = const()[name = string("op_3152_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_3152_cast_fp16 = transpose(perm = var_3152_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_cast_fp16 = reshape(shape = concat_107x, x = var_3152_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468209920)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_3185_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3185_cast_fp16")];
            int32 var_3183 = const()[name = string("op_3183"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_69_cast_fp16 = concat(axis = var_3183, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3185_cast_fp16))[name = string("doubled_69_cast_fp16")];
            tensor<int32, [1]> out_35_axes_0 = const()[name = string("out_35_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_35_gamma_0_to_fp16 = const()[name = string("out_35_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476598592)))];
            fp16 var_3195_to_fp16 = const()[name = string("op_3195_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_3195_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3206_split_sizes_0 = const()[name = string("op_3206_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3206_axis_0 = const()[name = string("op_3206_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3206_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3206_cast_fp16_1 = split(axis = var_3206_axis_0, split_sizes = var_3206_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3206_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476606848)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_8_mlp_gate_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_3223_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3223_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501772736)))];
            tensor<int32, [2]> var_3229_strides_0 = const()[name = string("op_3229_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3229_pad_type_0 = const()[name = string("op_3229_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3229_pad_0 = const()[name = string("op_3229_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3229_dilations_0 = const()[name = string("op_3229_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3229_groups_0 = const()[name = string("op_3229_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_3229_cast_fp16 = conv(dilations = var_3229_dilations_0, groups = var_3229_groups_0, pad = var_3229_pad_0, pad_type = var_3229_pad_type_0, strides = var_3229_strides_0, weight = layers_8_mlp_up_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("op_3229_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_cast_fp16 = mul(x = var_3223_cast_fp16, y = var_3229_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526938624)))];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_3247_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3247_cast_fp16")];
            int32 var_3245 = const()[name = string("op_3245"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_73_cast_fp16 = concat(axis = var_3245, interleave = doubled_73_interleave_0, values = (hidden_states_cast_fp16, var_3247_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_axes_0 = const()[name = string("out_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_gamma_0_to_fp16 = const()[name = string("out_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552104512)))];
            fp16 var_3257_to_fp16 = const()[name = string("op_3257_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3257_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3268_split_sizes_0 = const()[name = string("op_3268_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3268_axis_0 = const()[name = string("op_3268_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states, tensor<fp16, [1, 2048, 1, 16]> var_3268_cast_fp16_1 = split(axis = var_3268_axis_0, split_sizes = var_3268_split_sizes_0, x = out_cast_fp16)[name = string("op_3268_cast_fp16")];
        } -> (hidden_states);
    func length_32<ios18>(tensor<fp32, [1, 2048, 1, 32]> inputs_embeds, state<tensor<fp16, [9, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [32]> position_index_seed, state<tensor<fp16, [9, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725952))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17321280))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17308928))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17327488))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29922816))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29910464))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29929024))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42516160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42512000))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42518272))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243840))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243264))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47244160))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51442688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51438528))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444800))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64040128))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64027776))))[name = string("layers_1_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64046336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76633472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76629312))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76635584))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80834112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80829952))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80836224))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361152))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81360576))))[name = string("layers_2_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361472))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85560000))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85555840))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85562112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98145088))))[name = string("layers_2_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110758976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110746624))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110765184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123352320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123348160))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123354432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127552960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127548800))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127555072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128079424))))[name = string("layers_3_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080320))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140675648))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140663296))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140681856))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153277184))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153264832))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153283392))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165870528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165866368))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165872640))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170071168))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170067008))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170073280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170597632))))[name = string("layers_4_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174797056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174792896))))[name = string("layers_4_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174799168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187394496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187382144))))[name = string("layers_4_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187400704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199996032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199983680))))[name = string("layers_4_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_4_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200002240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212589376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212585216))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212591488))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216790016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216785856))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216792128))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317056))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217316480))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317376))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221515904))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221511744))))[name = string("layers_5_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221518016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234113344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234100992))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234119552))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246714880))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246702528))))[name = string("layers_5_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_5_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246721088))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259308224))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259304064))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259310336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263508864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263504704))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263510976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035328))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264036224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268234752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268230592))))[name = string("layers_6_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268236864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280832192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280819840))))[name = string("layers_6_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280838400))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293433728))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293421376))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293439936))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297638464))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297634304))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297640576))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165504))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298164928))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165824))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302364352))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302360192))))[name = string("layers_7_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302366464))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314961792))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314949440))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(32)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [32]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_280 = const()[name = string("op_280"), val = int32(0)];
            bool var_282_exclusive_0 = const()[name = string("op_282_exclusive_0"), val = bool(false)];
            bool var_282_reverse_0 = const()[name = string("op_282_reverse_0"), val = bool(false)];
            tensor<fp16, [32]> var_282_cast_fp16 = cumsum(axis = var_280, exclusive = var_282_exclusive_0, reverse = var_282_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_282_cast_fp16")];
            fp16 var_284_promoted_to_fp16 = const()[name = string("op_284_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [32]> position_offsets_cast_fp16 = sub(x = var_282_cast_fp16, y = var_284_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 32]> var_287_cast_fp16 = expand_dims(axes = var_287_axes_0, x = position_offsets_cast_fp16)[name = string("op_287_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_19")];
            tensor<fp16, [1, 32]> position_ids_1_cast_fp16 = add(x = var_287_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 32]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_18")];
            tensor<bool, [1, 32]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 32]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 32]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314968000)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_8_y_0 = const()[name = string("greater_equal_8_y_0"), val = int32(0)];
            tensor<bool, [1, 32]> greater_equal_8 = greater_equal(x = select_0, y = greater_equal_8_y_0)[name = string("greater_equal_8")];
            int32 slice_by_index_8 = const()[name = string("slice_by_index_8"), val = int32(32768)];
            tensor<int32, [1, 32]> add_8 = add(x = select_0, y = slice_by_index_8)[name = string("add_8")];
            tensor<int32, [1, 32]> select_8 = select(a = select_0, b = add_8, cond = greater_equal_8)[name = string("select_8")];
            int32 cos_1_cast_fp16_axis_4 = const()[name = string("cos_1_cast_fp16_axis_4"), val = int32(0)];
            tensor<fp16, [1, 32, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_4, batch_dims = cos_1_batch_dims_0, indices = select_8, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323356672)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_4 = const()[name = string("sin_1_cast_fp16_axis_4"), val = int32(0)];
            tensor<fp16, [1, 32, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_4, batch_dims = sin_1_batch_dims_0, indices = select_8, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_306_perm_0 = const()[name = string("op_306_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_308_axes_0 = const()[name = string("op_308_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 32]> var_306_cast_fp16 = transpose(perm = var_306_perm_0, x = cos_1_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 1, 128, 32]> var_308_cast_fp16 = expand_dims(axes = var_308_axes_0, x = var_306_cast_fp16)[name = string("op_308_cast_fp16")];
            tensor<int32, [3]> var_313_perm_0 = const()[name = string("op_313_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 32]> var_313_cast_fp16 = transpose(perm = var_313_perm_0, x = sin_1_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 1, 128, 32]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = var_313_cast_fp16)[name = string("op_315_cast_fp16")];
            tensor<int32, [1]> var_334_axes_0 = const()[name = string("op_334_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 32, 1]> var_334 = expand_dims(axes = var_334_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_334")];
            tensor<int32, [1, 1, 2048]> var_327 = const()[name = string("op_327"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331745344)))];
            tensor<bool, [1, 32, 2048]> var_335 = greater(x = var_327, y = var_334)[name = string("op_335")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_342_axes_0 = const()[name = string("op_342_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 32, 2048]> var_335_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_335)[name = string("cast_17")];
            tensor<fp16, [1, 1, 32, 2048]> var_342_cast_fp16 = expand_dims(axes = var_342_axes_0, x = var_335_to_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_346_promoted_to_fp16 = const()[name = string("op_346_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 32]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_342_cast_fp16)[name = string("transpose_147")];
            tensor<bool, [1, 1, 2048, 32]> var_347_cast_fp16 = equal(x = mask_cast_fp16, y = var_346_promoted_to_fp16)[name = string("op_347_cast_fp16")];
            fp16 var_348_to_fp16 = const()[name = string("op_348_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 32]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_348_to_fp16, cond = var_347_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_16")];
            tensor<fp16, [1, 2048, 1, 32]> var_358_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_358_cast_fp16")];
            int32 var_356 = const()[name = string("op_356"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_1_cast_fp16 = concat(axis = var_356, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_358_cast_fp16))[name = string("doubled_1_cast_fp16")];
            tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_1_gamma_0_to_fp16 = const()[name = string("out_1_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331753600)))];
            fp16 var_368_to_fp16 = const()[name = string("op_368_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_368_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_379_split_sizes_0 = const()[name = string("op_379_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_379_axis_0 = const()[name = string("op_379_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_379_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_379_cast_fp16_1 = split(axis = var_379_axis_0, split_sizes = var_379_split_sizes_0, x = out_1_cast_fp16)[name = string("op_379_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331761856)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_379_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_436_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_436_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_443_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_443_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_447_cast_fp16 = mul(x = x_1_cast_fp16, y = var_308_cast_fp16)[name = string("op_447_cast_fp16")];
            tensor<int32, [2]> var_448_split_sizes_0 = const()[name = string("op_448_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_448_axis_0 = const()[name = string("op_448_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_448_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_448_cast_fp16_1 = split(axis = var_448_axis_0, split_sizes = var_448_split_sizes_0, x = x_1_cast_fp16)[name = string("op_448_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_450_cast_fp16 = mul(x = var_448_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_450_cast_fp16")];
            int32 var_452 = const()[name = string("op_452"), val = int32(-2)];
            bool var_453_interleave_0 = const()[name = string("op_453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_453_cast_fp16 = concat(axis = var_452, interleave = var_453_interleave_0, values = (var_450_cast_fp16, var_448_cast_fp16_0))[name = string("op_453_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_454_cast_fp16 = mul(x = var_453_cast_fp16, y = var_315_cast_fp16)[name = string("op_454_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_3_cast_fp16 = add(x = var_447_cast_fp16, y = var_454_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_460_cast_fp16 = mul(x = var_436_cast_fp16, y = var_308_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [2]> var_461_split_sizes_0 = const()[name = string("op_461_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_461_axis_0 = const()[name = string("op_461_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_461_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_461_cast_fp16_1 = split(axis = var_461_axis_0, split_sizes = var_461_split_sizes_0, x = var_436_cast_fp16)[name = string("op_461_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_463_cast_fp16 = mul(x = var_461_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_463_cast_fp16")];
            int32 var_465 = const()[name = string("op_465"), val = int32(-2)];
            bool var_466_interleave_0 = const()[name = string("op_466_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_466_cast_fp16 = concat(axis = var_465, interleave = var_466_interleave_0, values = (var_463_cast_fp16, var_461_cast_fp16_0))[name = string("op_466_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_467_cast_fp16 = mul(x = var_466_cast_fp16, y = var_315_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_5_cast_fp16 = add(x = var_460_cast_fp16, y = var_467_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_72_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_72 = read_state(input = key_cache)[name = string("coreml_update_state_72")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_443_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_73_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_73 = read_state(input = value_cache)[name = string("coreml_update_state_73")];
            tensor<int32, [4]> var_537_begin_0 = const()[name = string("op_537_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_537_end_0 = const()[name = string("op_537_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_537_end_mask_0 = const()[name = string("op_537_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_537_cast_fp16 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = coreml_update_state_72)[name = string("op_537_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_540_axis_0 = const()[name = string("op_540_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_1 = split(axis = var_540_axis_0, split_sizes = tile_0, x = var_537_cast_fp16)[name = string("op_540_cast_fp16")];
            tensor<int32, [4]> var_547_begin_0 = const()[name = string("op_547_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_547_end_0 = const()[name = string("op_547_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_547_end_mask_0 = const()[name = string("op_547_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_547_cast_fp16 = slice_by_index(begin = var_547_begin_0, end = var_547_end_0, end_mask = var_547_end_mask_0, x = coreml_update_state_73)[name = string("op_547_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_550_axis_0 = const()[name = string("op_550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_1 = split(axis = var_550_axis_0, split_sizes = tile_1, x = var_547_cast_fp16)[name = string("op_550_cast_fp16")];
            tensor<int32, [2]> var_553_split_sizes_0 = const()[name = string("op_553_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_553_axis_0 = const()[name = string("op_553_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_553_0, tensor<fp16, [1, 8, 128, 32]> var_553_1 = split(axis = var_553_axis_0, split_sizes = var_553_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_553")];
            bool attn_weights_1_transpose_x_0 = const()[name = string("attn_weights_1_transpose_x_0"), val = bool(false)];
            bool attn_weights_1_transpose_y_0 = const()[name = string("attn_weights_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_540_cast_fp16_0, y = var_553_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_556_to_fp16 = const()[name = string("op_556_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_556_to_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_5_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
            int32 var_560 = const()[name = string("op_560"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_7_cast_fp16 = softmax(axis = var_560, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_566_transpose_x_1 = const()[name = string("op_566_transpose_x_1"), val = bool(true)];
            bool var_566_transpose_y_1 = const()[name = string("op_566_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_566_cast_fp16 = matmul(transpose_x = var_566_transpose_x_1, transpose_y = var_566_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_550_cast_fp16_0)[name = string("op_566_cast_fp16")];
            bool attn_weights_9_transpose_x_0 = const()[name = string("attn_weights_9_transpose_x_0"), val = bool(false)];
            bool attn_weights_9_transpose_y_0 = const()[name = string("attn_weights_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_9_cast_fp16 = matmul(transpose_x = attn_weights_9_transpose_x_0, transpose_y = attn_weights_9_transpose_y_0, x = var_540_cast_fp16_1, y = var_553_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_568_to_fp16 = const()[name = string("op_568_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_11_cast_fp16 = mul(x = attn_weights_9_cast_fp16, y = var_568_to_fp16)[name = string("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_13_cast_fp16 = add(x = attn_weights_11_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_13_cast_fp16")];
            int32 var_572 = const()[name = string("op_572"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_15_cast_fp16 = softmax(axis = var_572, x = attn_weights_13_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(true)];
            bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_15_cast_fp16, y = var_550_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_580 = const()[name = string("op_580"), val = int32(1)];
            bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_3_cast_fp16 = concat(axis = var_580, interleave = attn_output_3_interleave_0, values = (var_566_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_584_perm_0 = const()[name = string("op_584_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_584_cast_fp16 = transpose(perm = var_584_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_584_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332810496)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_617_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_617_cast_fp16")];
            int32 var_615 = const()[name = string("op_615"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_5_cast_fp16 = concat(axis = var_615, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_617_cast_fp16))[name = string("doubled_5_cast_fp16")];
            tensor<int32, [1]> out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_3_gamma_0_to_fp16 = const()[name = string("out_3_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341199168)))];
            fp16 var_627_to_fp16 = const()[name = string("op_627_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_627_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_638_split_sizes_0 = const()[name = string("op_638_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_638_axis_0 = const()[name = string("op_638_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_638_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_638_cast_fp16_1 = split(axis = var_638_axis_0, split_sizes = var_638_split_sizes_0, x = out_3_cast_fp16)[name = string("op_638_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_655_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_655_cast_fp16")];
            tensor<int32, [2]> var_661_strides_0 = const()[name = string("op_661_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_661_pad_type_0 = const()[name = string("op_661_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_661_pad_0 = const()[name = string("op_661_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_661_dilations_0 = const()[name = string("op_661_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_661_groups_0 = const()[name = string("op_661_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_661_cast_fp16 = conv(dilations = var_661_dilations_0, groups = var_661_groups_0, pad = var_661_pad_0, pad_type = var_661_pad_type_0, strides = var_661_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("op_661_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_9_cast_fp16 = mul(x = var_655_cast_fp16, y = var_661_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_679_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_679_cast_fp16")];
            int32 var_677 = const()[name = string("op_677"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_9_cast_fp16 = concat(axis = var_677, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_679_cast_fp16))[name = string("doubled_9_cast_fp16")];
            tensor<int32, [1]> out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_5_gamma_0_to_fp16 = const()[name = string("out_5_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341207424)))];
            fp16 var_689_to_fp16 = const()[name = string("op_689_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_689_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_700_split_sizes_0 = const()[name = string("op_700_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_700_axis_0 = const()[name = string("op_700_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_700_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_700_cast_fp16_1 = split(axis = var_700_axis_0, split_sizes = var_700_split_sizes_0, x = out_5_cast_fp16)[name = string("op_700_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341215680)))];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = var_700_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_757_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_757_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_764_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_764_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_768_cast_fp16 = mul(x = x_11_cast_fp16, y = var_308_cast_fp16)[name = string("op_768_cast_fp16")];
            tensor<int32, [2]> var_769_split_sizes_0 = const()[name = string("op_769_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_769_axis_0 = const()[name = string("op_769_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_769_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_769_cast_fp16_1 = split(axis = var_769_axis_0, split_sizes = var_769_split_sizes_0, x = x_11_cast_fp16)[name = string("op_769_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_771_cast_fp16 = mul(x = var_769_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_771_cast_fp16")];
            int32 var_773 = const()[name = string("op_773"), val = int32(-2)];
            bool var_774_interleave_0 = const()[name = string("op_774_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_774_cast_fp16 = concat(axis = var_773, interleave = var_774_interleave_0, values = (var_771_cast_fp16, var_769_cast_fp16_0))[name = string("op_774_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_775_cast_fp16 = mul(x = var_774_cast_fp16, y = var_315_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_9_cast_fp16 = add(x = var_768_cast_fp16, y = var_775_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_781_cast_fp16 = mul(x = var_757_cast_fp16, y = var_308_cast_fp16)[name = string("op_781_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_782_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = var_757_cast_fp16)[name = string("op_782_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_784_cast_fp16 = mul(x = var_782_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_784_cast_fp16")];
            int32 var_786 = const()[name = string("op_786"), val = int32(-2)];
            bool var_787_interleave_0 = const()[name = string("op_787_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_787_cast_fp16 = concat(axis = var_786, interleave = var_787_interleave_0, values = (var_784_cast_fp16, var_782_cast_fp16_0))[name = string("op_787_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_788_cast_fp16 = mul(x = var_787_cast_fp16, y = var_315_cast_fp16)[name = string("op_788_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_15_cast_fp16 = add(x = var_781_cast_fp16, y = var_788_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_72)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_74_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_74 = read_state(input = key_cache)[name = string("coreml_update_state_74")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_764_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_73)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_75_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_75 = read_state(input = value_cache)[name = string("coreml_update_state_75")];
            tensor<int32, [4]> var_858_begin_0 = const()[name = string("op_858_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_858_end_0 = const()[name = string("op_858_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_858_end_mask_0 = const()[name = string("op_858_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_858_cast_fp16 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = coreml_update_state_74)[name = string("op_858_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_861_axis_0 = const()[name = string("op_861_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_1 = split(axis = var_861_axis_0, split_sizes = tile_2, x = var_858_cast_fp16)[name = string("op_861_cast_fp16")];
            tensor<int32, [4]> var_868_begin_0 = const()[name = string("op_868_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_868_end_0 = const()[name = string("op_868_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_868_end_mask_0 = const()[name = string("op_868_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_868_cast_fp16 = slice_by_index(begin = var_868_begin_0, end = var_868_end_0, end_mask = var_868_end_mask_0, x = coreml_update_state_75)[name = string("op_868_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_871_axis_0 = const()[name = string("op_871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_1 = split(axis = var_871_axis_0, split_sizes = tile_3, x = var_868_cast_fp16)[name = string("op_871_cast_fp16")];
            tensor<int32, [2]> var_874_split_sizes_0 = const()[name = string("op_874_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_874_axis_0 = const()[name = string("op_874_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_874_0, tensor<fp16, [1, 8, 128, 32]> var_874_1 = split(axis = var_874_axis_0, split_sizes = var_874_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_874")];
            bool attn_weights_17_transpose_x_0 = const()[name = string("attn_weights_17_transpose_x_0"), val = bool(false)];
            bool attn_weights_17_transpose_y_0 = const()[name = string("attn_weights_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_17_cast_fp16 = matmul(transpose_x = attn_weights_17_transpose_x_0, transpose_y = attn_weights_17_transpose_y_0, x = var_861_cast_fp16_0, y = var_874_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_877_to_fp16 = const()[name = string("op_877_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_19_cast_fp16 = mul(x = attn_weights_17_cast_fp16, y = var_877_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            int32 var_881 = const()[name = string("op_881"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_23_cast_fp16 = softmax(axis = var_881, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_887_transpose_x_1 = const()[name = string("op_887_transpose_x_1"), val = bool(true)];
            bool var_887_transpose_y_1 = const()[name = string("op_887_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_887_cast_fp16 = matmul(transpose_x = var_887_transpose_x_1, transpose_y = var_887_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_871_cast_fp16_0)[name = string("op_887_cast_fp16")];
            bool attn_weights_25_transpose_x_0 = const()[name = string("attn_weights_25_transpose_x_0"), val = bool(false)];
            bool attn_weights_25_transpose_y_0 = const()[name = string("attn_weights_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_0, transpose_y = attn_weights_25_transpose_y_0, x = var_861_cast_fp16_1, y = var_874_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_889_to_fp16 = const()[name = string("op_889_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_889_to_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_29_cast_fp16 = add(x = attn_weights_27_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
            int32 var_893 = const()[name = string("op_893"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_31_cast_fp16 = softmax(axis = var_893, x = attn_weights_29_cast_fp16)[name = string("attn_weights_31_cast_fp16")];
            bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(true)];
            bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_31_cast_fp16, y = var_871_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_901 = const()[name = string("op_901"), val = int32(1)];
            bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_11_cast_fp16 = concat(axis = var_901, interleave = attn_output_11_interleave_0, values = (var_887_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_905_perm_0 = const()[name = string("op_905_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_905_cast_fp16 = transpose(perm = var_905_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_905_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_938_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_938_cast_fp16")];
            int32 var_936 = const()[name = string("op_936"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_13_cast_fp16 = concat(axis = var_936, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_938_cast_fp16))[name = string("doubled_13_cast_fp16")];
            tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_7_gamma_0_to_fp16 = const()[name = string("out_7_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342264320)))];
            fp16 var_948_to_fp16 = const()[name = string("op_948_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_948_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_959_split_sizes_0 = const()[name = string("op_959_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_959_axis_0 = const()[name = string("op_959_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_959_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_959_cast_fp16_1 = split(axis = var_959_axis_0, split_sizes = var_959_split_sizes_0, x = out_7_cast_fp16)[name = string("op_959_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_959_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_976_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_976_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342272576)))];
            tensor<int32, [2]> var_982_strides_0 = const()[name = string("op_982_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_982_pad_type_0 = const()[name = string("op_982_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_982_pad_0 = const()[name = string("op_982_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_982_dilations_0 = const()[name = string("op_982_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_982_groups_0 = const()[name = string("op_982_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_982_cast_fp16 = conv(dilations = var_982_dilations_0, groups = var_982_groups_0, pad = var_982_pad_0, pad_type = var_982_pad_type_0, strides = var_982_strides_0, weight = layers_1_mlp_up_proj_weight_to_fp16, x = var_959_cast_fp16_0)[name = string("op_982_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_19_cast_fp16 = mul(x = var_976_cast_fp16, y = var_982_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1000_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1000_cast_fp16")];
            int32 var_998 = const()[name = string("op_998"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_17_cast_fp16 = concat(axis = var_998, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1000_cast_fp16))[name = string("doubled_17_cast_fp16")];
            tensor<int32, [1]> out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_9_gamma_0_to_fp16 = const()[name = string("out_9_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367438464)))];
            fp16 var_1010_to_fp16 = const()[name = string("op_1010_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1010_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1021_split_sizes_0 = const()[name = string("op_1021_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1021_axis_0 = const()[name = string("op_1021_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1021_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1021_cast_fp16_1 = split(axis = var_1021_axis_0, split_sizes = var_1021_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1021_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367446720)))];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = var_1021_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_1078_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1078_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_1085_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1085_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1089_cast_fp16 = mul(x = x_21_cast_fp16, y = var_308_cast_fp16)[name = string("op_1089_cast_fp16")];
            tensor<int32, [2]> var_1090_split_sizes_0 = const()[name = string("op_1090_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1090_axis_0 = const()[name = string("op_1090_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_1090_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_1090_cast_fp16_1 = split(axis = var_1090_axis_0, split_sizes = var_1090_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1090_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_1092_cast_fp16 = mul(x = var_1090_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            int32 var_1094 = const()[name = string("op_1094"), val = int32(-2)];
            bool var_1095_interleave_0 = const()[name = string("op_1095_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_1095_cast_fp16 = concat(axis = var_1094, interleave = var_1095_interleave_0, values = (var_1092_cast_fp16, var_1090_cast_fp16_0))[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1096_cast_fp16 = mul(x = var_1095_cast_fp16, y = var_315_cast_fp16)[name = string("op_1096_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_15_cast_fp16 = add(x = var_1089_cast_fp16, y = var_1096_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1102_cast_fp16 = mul(x = var_1078_cast_fp16, y = var_308_cast_fp16)[name = string("op_1102_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_1103_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = var_1078_cast_fp16)[name = string("op_1103_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_1105_cast_fp16 = mul(x = var_1103_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1105_cast_fp16")];
            int32 var_1107 = const()[name = string("op_1107"), val = int32(-2)];
            bool var_1108_interleave_0 = const()[name = string("op_1108_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_1108_cast_fp16 = concat(axis = var_1107, interleave = var_1108_interleave_0, values = (var_1105_cast_fp16, var_1103_cast_fp16_0))[name = string("op_1108_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1109_cast_fp16 = mul(x = var_1108_cast_fp16, y = var_315_cast_fp16)[name = string("op_1109_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_25_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1109_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_74)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_76_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_76 = read_state(input = key_cache)[name = string("coreml_update_state_76")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1085_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_75)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_77_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_77 = read_state(input = value_cache)[name = string("coreml_update_state_77")];
            tensor<int32, [4]> var_1179_begin_0 = const()[name = string("op_1179_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1179_end_0 = const()[name = string("op_1179_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1179_end_mask_0 = const()[name = string("op_1179_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1179_cast_fp16 = slice_by_index(begin = var_1179_begin_0, end = var_1179_end_0, end_mask = var_1179_end_mask_0, x = coreml_update_state_76)[name = string("op_1179_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1182_axis_0 = const()[name = string("op_1182_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_1 = split(axis = var_1182_axis_0, split_sizes = tile_4, x = var_1179_cast_fp16)[name = string("op_1182_cast_fp16")];
            tensor<int32, [4]> var_1189_begin_0 = const()[name = string("op_1189_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1189_end_0 = const()[name = string("op_1189_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1189_end_mask_0 = const()[name = string("op_1189_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1189_cast_fp16 = slice_by_index(begin = var_1189_begin_0, end = var_1189_end_0, end_mask = var_1189_end_mask_0, x = coreml_update_state_77)[name = string("op_1189_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1192_axis_0 = const()[name = string("op_1192_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_1 = split(axis = var_1192_axis_0, split_sizes = tile_5, x = var_1189_cast_fp16)[name = string("op_1192_cast_fp16")];
            tensor<int32, [2]> var_1195_split_sizes_0 = const()[name = string("op_1195_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1195_axis_0 = const()[name = string("op_1195_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1195_0, tensor<fp16, [1, 8, 128, 32]> var_1195_1 = split(axis = var_1195_axis_0, split_sizes = var_1195_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1195")];
            bool attn_weights_33_transpose_x_0 = const()[name = string("attn_weights_33_transpose_x_0"), val = bool(false)];
            bool attn_weights_33_transpose_y_0 = const()[name = string("attn_weights_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_33_cast_fp16 = matmul(transpose_x = attn_weights_33_transpose_x_0, transpose_y = attn_weights_33_transpose_y_0, x = var_1182_cast_fp16_0, y = var_1195_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1198_to_fp16 = const()[name = string("op_1198_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_35_cast_fp16 = mul(x = attn_weights_33_cast_fp16, y = var_1198_to_fp16)[name = string("attn_weights_35_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_37_cast_fp16 = add(x = attn_weights_35_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_37_cast_fp16")];
            int32 var_1202 = const()[name = string("op_1202"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_39_cast_fp16 = softmax(axis = var_1202, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1208_transpose_x_1 = const()[name = string("op_1208_transpose_x_1"), val = bool(true)];
            bool var_1208_transpose_y_1 = const()[name = string("op_1208_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1208_cast_fp16 = matmul(transpose_x = var_1208_transpose_x_1, transpose_y = var_1208_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1192_cast_fp16_0)[name = string("op_1208_cast_fp16")];
            bool attn_weights_41_transpose_x_0 = const()[name = string("attn_weights_41_transpose_x_0"), val = bool(false)];
            bool attn_weights_41_transpose_y_0 = const()[name = string("attn_weights_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_41_cast_fp16 = matmul(transpose_x = attn_weights_41_transpose_x_0, transpose_y = attn_weights_41_transpose_y_0, x = var_1182_cast_fp16_1, y = var_1195_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1210_to_fp16 = const()[name = string("op_1210_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_43_cast_fp16 = mul(x = attn_weights_41_cast_fp16, y = var_1210_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            int32 var_1214 = const()[name = string("op_1214"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_47_cast_fp16 = softmax(axis = var_1214, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")];
            bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(true)];
            bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_47_cast_fp16, y = var_1192_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1222 = const()[name = string("op_1222"), val = int32(1)];
            bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_19_cast_fp16 = concat(axis = var_1222, interleave = attn_output_19_interleave_0, values = (var_1208_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1226_perm_0 = const()[name = string("op_1226_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_1226_cast_fp16 = transpose(perm = var_1226_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1226_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1259_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            int32 var_1257 = const()[name = string("op_1257"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_21_cast_fp16 = concat(axis = var_1257, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1259_cast_fp16))[name = string("doubled_21_cast_fp16")];
            tensor<int32, [1]> out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_11_gamma_0_to_fp16 = const()[name = string("out_11_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368495360)))];
            fp16 var_1269_to_fp16 = const()[name = string("op_1269_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1269_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1280_split_sizes_0 = const()[name = string("op_1280_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1280_axis_0 = const()[name = string("op_1280_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1280_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1280_cast_fp16_1 = split(axis = var_1280_axis_0, split_sizes = var_1280_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1280_cast_fp16")];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_1297_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1297_cast_fp16")];
            tensor<int32, [2]> var_1303_strides_0 = const()[name = string("op_1303_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1303_pad_type_0 = const()[name = string("op_1303_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1303_pad_0 = const()[name = string("op_1303_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1303_dilations_0 = const()[name = string("op_1303_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1303_groups_0 = const()[name = string("op_1303_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_1303_cast_fp16 = conv(dilations = var_1303_dilations_0, groups = var_1303_groups_0, pad = var_1303_pad_0, pad_type = var_1303_pad_type_0, strides = var_1303_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("op_1303_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_29_cast_fp16 = mul(x = var_1297_cast_fp16, y = var_1303_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1321_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1321_cast_fp16")];
            int32 var_1319 = const()[name = string("op_1319"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_25_cast_fp16 = concat(axis = var_1319, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1321_cast_fp16))[name = string("doubled_25_cast_fp16")];
            tensor<int32, [1]> out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_13_gamma_0_to_fp16 = const()[name = string("out_13_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368503616)))];
            fp16 var_1331_to_fp16 = const()[name = string("op_1331_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1331_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1342_split_sizes_0 = const()[name = string("op_1342_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1342_axis_0 = const()[name = string("op_1342_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1342_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1342_cast_fp16_1 = split(axis = var_1342_axis_0, split_sizes = var_1342_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1342_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368511872)))];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = var_1342_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_1399_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_1406_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1406_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1410_cast_fp16 = mul(x = x_31_cast_fp16, y = var_308_cast_fp16)[name = string("op_1410_cast_fp16")];
            tensor<int32, [2]> var_1411_split_sizes_0 = const()[name = string("op_1411_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1411_axis_0 = const()[name = string("op_1411_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_1411_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_1411_cast_fp16_1 = split(axis = var_1411_axis_0, split_sizes = var_1411_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1411_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_1413_cast_fp16 = mul(x = var_1411_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1413_cast_fp16")];
            int32 var_1415 = const()[name = string("op_1415"), val = int32(-2)];
            bool var_1416_interleave_0 = const()[name = string("op_1416_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_1416_cast_fp16 = concat(axis = var_1415, interleave = var_1416_interleave_0, values = (var_1413_cast_fp16, var_1411_cast_fp16_0))[name = string("op_1416_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1417_cast_fp16 = mul(x = var_1416_cast_fp16, y = var_315_cast_fp16)[name = string("op_1417_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_21_cast_fp16 = add(x = var_1410_cast_fp16, y = var_1417_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1423_cast_fp16 = mul(x = var_1399_cast_fp16, y = var_308_cast_fp16)[name = string("op_1423_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_1424_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = var_1399_cast_fp16)[name = string("op_1424_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_1426_cast_fp16 = mul(x = var_1424_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1426_cast_fp16")];
            int32 var_1428 = const()[name = string("op_1428"), val = int32(-2)];
            bool var_1429_interleave_0 = const()[name = string("op_1429_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_1429_cast_fp16 = concat(axis = var_1428, interleave = var_1429_interleave_0, values = (var_1426_cast_fp16, var_1424_cast_fp16_0))[name = string("op_1429_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1430_cast_fp16 = mul(x = var_1429_cast_fp16, y = var_315_cast_fp16)[name = string("op_1430_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_35_cast_fp16 = add(x = var_1423_cast_fp16, y = var_1430_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_76)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_78_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_78 = read_state(input = key_cache)[name = string("coreml_update_state_78")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1406_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_77)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_79_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_79 = read_state(input = value_cache)[name = string("coreml_update_state_79")];
            tensor<int32, [4]> var_1500_begin_0 = const()[name = string("op_1500_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1500_end_0 = const()[name = string("op_1500_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1500_end_mask_0 = const()[name = string("op_1500_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1500_cast_fp16 = slice_by_index(begin = var_1500_begin_0, end = var_1500_end_0, end_mask = var_1500_end_mask_0, x = coreml_update_state_78)[name = string("op_1500_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1503_axis_0 = const()[name = string("op_1503_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_1 = split(axis = var_1503_axis_0, split_sizes = tile_6, x = var_1500_cast_fp16)[name = string("op_1503_cast_fp16")];
            tensor<int32, [4]> var_1510_begin_0 = const()[name = string("op_1510_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1510_end_0 = const()[name = string("op_1510_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1510_end_mask_0 = const()[name = string("op_1510_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1510_cast_fp16 = slice_by_index(begin = var_1510_begin_0, end = var_1510_end_0, end_mask = var_1510_end_mask_0, x = coreml_update_state_79)[name = string("op_1510_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1513_axis_0 = const()[name = string("op_1513_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_1 = split(axis = var_1513_axis_0, split_sizes = tile_7, x = var_1510_cast_fp16)[name = string("op_1513_cast_fp16")];
            tensor<int32, [2]> var_1516_split_sizes_0 = const()[name = string("op_1516_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1516_axis_0 = const()[name = string("op_1516_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1516_0, tensor<fp16, [1, 8, 128, 32]> var_1516_1 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1516")];
            bool attn_weights_49_transpose_x_0 = const()[name = string("attn_weights_49_transpose_x_0"), val = bool(false)];
            bool attn_weights_49_transpose_y_0 = const()[name = string("attn_weights_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_0, transpose_y = attn_weights_49_transpose_y_0, x = var_1503_cast_fp16_0, y = var_1516_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_1519_to_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_53_cast_fp16 = add(x = attn_weights_51_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_53_cast_fp16")];
            int32 var_1523 = const()[name = string("op_1523"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_55_cast_fp16 = softmax(axis = var_1523, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1529_transpose_x_1 = const()[name = string("op_1529_transpose_x_1"), val = bool(true)];
            bool var_1529_transpose_y_1 = const()[name = string("op_1529_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1529_cast_fp16 = matmul(transpose_x = var_1529_transpose_x_1, transpose_y = var_1529_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1513_cast_fp16_0)[name = string("op_1529_cast_fp16")];
            bool attn_weights_57_transpose_x_0 = const()[name = string("attn_weights_57_transpose_x_0"), val = bool(false)];
            bool attn_weights_57_transpose_y_0 = const()[name = string("attn_weights_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_57_cast_fp16 = matmul(transpose_x = attn_weights_57_transpose_x_0, transpose_y = attn_weights_57_transpose_y_0, x = var_1503_cast_fp16_1, y = var_1516_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1531_to_fp16 = const()[name = string("op_1531_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_59_cast_fp16 = mul(x = attn_weights_57_cast_fp16, y = var_1531_to_fp16)[name = string("attn_weights_59_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_61_cast_fp16 = add(x = attn_weights_59_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_61_cast_fp16")];
            int32 var_1535 = const()[name = string("op_1535"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_63_cast_fp16 = softmax(axis = var_1535, x = attn_weights_61_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(true)];
            bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_63_cast_fp16, y = var_1513_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1543 = const()[name = string("op_1543"), val = int32(1)];
            bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_27_cast_fp16 = concat(axis = var_1543, interleave = attn_output_27_interleave_0, values = (var_1529_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1547_perm_0 = const()[name = string("op_1547_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_1547_cast_fp16 = transpose(perm = var_1547_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1547_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369560512)))];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1580_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1580_cast_fp16")];
            int32 var_1578 = const()[name = string("op_1578"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_29_cast_fp16 = concat(axis = var_1578, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1580_cast_fp16))[name = string("doubled_29_cast_fp16")];
            tensor<int32, [1]> out_15_axes_0 = const()[name = string("out_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_15_gamma_0_to_fp16 = const()[name = string("out_15_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377949184)))];
            fp16 var_1590_to_fp16 = const()[name = string("op_1590_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_1590_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1601_split_sizes_0 = const()[name = string("op_1601_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1601_axis_0 = const()[name = string("op_1601_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1601_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1601_cast_fp16_1 = split(axis = var_1601_axis_0, split_sizes = var_1601_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1601_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_1618_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1618_cast_fp16")];
            tensor<int32, [2]> var_1624_strides_0 = const()[name = string("op_1624_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1624_pad_type_0 = const()[name = string("op_1624_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1624_pad_0 = const()[name = string("op_1624_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1624_dilations_0 = const()[name = string("op_1624_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1624_groups_0 = const()[name = string("op_1624_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_1624_cast_fp16 = conv(dilations = var_1624_dilations_0, groups = var_1624_groups_0, pad = var_1624_pad_0, pad_type = var_1624_pad_type_0, strides = var_1624_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("op_1624_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_39_cast_fp16 = mul(x = var_1618_cast_fp16, y = var_1624_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1642_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1642_cast_fp16")];
            int32 var_1640 = const()[name = string("op_1640"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_33_cast_fp16 = concat(axis = var_1640, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1642_cast_fp16))[name = string("doubled_33_cast_fp16")];
            tensor<int32, [1]> out_17_axes_0 = const()[name = string("out_17_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_17_gamma_0_to_fp16 = const()[name = string("out_17_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377957440)))];
            fp16 var_1652_to_fp16 = const()[name = string("op_1652_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1652_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1663_split_sizes_0 = const()[name = string("op_1663_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1663_axis_0 = const()[name = string("op_1663_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1663_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1663_cast_fp16_1 = split(axis = var_1663_axis_0, split_sizes = var_1663_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1663_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377965696)))];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = var_1663_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_1720_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1720_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_1727_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1731_cast_fp16 = mul(x = x_41_cast_fp16, y = var_308_cast_fp16)[name = string("op_1731_cast_fp16")];
            tensor<int32, [2]> var_1732_split_sizes_0 = const()[name = string("op_1732_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1732_axis_0 = const()[name = string("op_1732_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_1732_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_1732_cast_fp16_1 = split(axis = var_1732_axis_0, split_sizes = var_1732_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1732_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_1734_cast_fp16 = mul(x = var_1732_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1734_cast_fp16")];
            int32 var_1736 = const()[name = string("op_1736"), val = int32(-2)];
            bool var_1737_interleave_0 = const()[name = string("op_1737_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_1737_cast_fp16 = concat(axis = var_1736, interleave = var_1737_interleave_0, values = (var_1734_cast_fp16, var_1732_cast_fp16_0))[name = string("op_1737_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1738_cast_fp16 = mul(x = var_1737_cast_fp16, y = var_315_cast_fp16)[name = string("op_1738_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_27_cast_fp16 = add(x = var_1731_cast_fp16, y = var_1738_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1744_cast_fp16 = mul(x = var_1720_cast_fp16, y = var_308_cast_fp16)[name = string("op_1744_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_1745_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = var_1720_cast_fp16)[name = string("op_1745_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_1747_cast_fp16 = mul(x = var_1745_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1747_cast_fp16")];
            int32 var_1749 = const()[name = string("op_1749"), val = int32(-2)];
            bool var_1750_interleave_0 = const()[name = string("op_1750_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_1750_cast_fp16 = concat(axis = var_1749, interleave = var_1750_interleave_0, values = (var_1747_cast_fp16, var_1745_cast_fp16_0))[name = string("op_1750_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1751_cast_fp16 = mul(x = var_1750_cast_fp16, y = var_315_cast_fp16)[name = string("op_1751_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_45_cast_fp16 = add(x = var_1744_cast_fp16, y = var_1751_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_78)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_80_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_80 = read_state(input = key_cache)[name = string("coreml_update_state_80")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1727_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_79)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_81_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_81 = read_state(input = value_cache)[name = string("coreml_update_state_81")];
            tensor<int32, [4]> var_1821_begin_0 = const()[name = string("op_1821_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1821_end_0 = const()[name = string("op_1821_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1821_end_mask_0 = const()[name = string("op_1821_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1821_cast_fp16 = slice_by_index(begin = var_1821_begin_0, end = var_1821_end_0, end_mask = var_1821_end_mask_0, x = coreml_update_state_80)[name = string("op_1821_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1824_axis_0 = const()[name = string("op_1824_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_1 = split(axis = var_1824_axis_0, split_sizes = tile_8, x = var_1821_cast_fp16)[name = string("op_1824_cast_fp16")];
            tensor<int32, [4]> var_1831_begin_0 = const()[name = string("op_1831_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1831_end_0 = const()[name = string("op_1831_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1831_end_mask_0 = const()[name = string("op_1831_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1831_cast_fp16 = slice_by_index(begin = var_1831_begin_0, end = var_1831_end_0, end_mask = var_1831_end_mask_0, x = coreml_update_state_81)[name = string("op_1831_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1834_axis_0 = const()[name = string("op_1834_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_1 = split(axis = var_1834_axis_0, split_sizes = tile_9, x = var_1831_cast_fp16)[name = string("op_1834_cast_fp16")];
            tensor<int32, [2]> var_1837_split_sizes_0 = const()[name = string("op_1837_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1837_axis_0 = const()[name = string("op_1837_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1837_0, tensor<fp16, [1, 8, 128, 32]> var_1837_1 = split(axis = var_1837_axis_0, split_sizes = var_1837_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1837")];
            bool attn_weights_65_transpose_x_0 = const()[name = string("attn_weights_65_transpose_x_0"), val = bool(false)];
            bool attn_weights_65_transpose_y_0 = const()[name = string("attn_weights_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_65_cast_fp16 = matmul(transpose_x = attn_weights_65_transpose_x_0, transpose_y = attn_weights_65_transpose_y_0, x = var_1824_cast_fp16_0, y = var_1837_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1840_to_fp16 = const()[name = string("op_1840_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_67_cast_fp16 = mul(x = attn_weights_65_cast_fp16, y = var_1840_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            int32 var_1844 = const()[name = string("op_1844"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_71_cast_fp16 = softmax(axis = var_1844, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1850_transpose_x_1 = const()[name = string("op_1850_transpose_x_1"), val = bool(true)];
            bool var_1850_transpose_y_1 = const()[name = string("op_1850_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1850_cast_fp16 = matmul(transpose_x = var_1850_transpose_x_1, transpose_y = var_1850_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1834_cast_fp16_0)[name = string("op_1850_cast_fp16")];
            bool attn_weights_73_transpose_x_0 = const()[name = string("attn_weights_73_transpose_x_0"), val = bool(false)];
            bool attn_weights_73_transpose_y_0 = const()[name = string("attn_weights_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_0, transpose_y = attn_weights_73_transpose_y_0, x = var_1824_cast_fp16_1, y = var_1837_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1852_to_fp16 = const()[name = string("op_1852_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1852_to_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_77_cast_fp16 = add(x = attn_weights_75_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_77_cast_fp16")];
            int32 var_1856 = const()[name = string("op_1856"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_79_cast_fp16 = softmax(axis = var_1856, x = attn_weights_77_cast_fp16)[name = string("attn_weights_79_cast_fp16")];
            bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(true)];
            bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_79_cast_fp16, y = var_1834_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1864 = const()[name = string("op_1864"), val = int32(1)];
            bool attn_output_35_interleave_0 = const()[name = string("attn_output_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_35_cast_fp16 = concat(axis = var_1864, interleave = attn_output_35_interleave_0, values = (var_1850_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1868_perm_0 = const()[name = string("op_1868_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_1868_cast_fp16 = transpose(perm = var_1868_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1868_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1901_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1901_cast_fp16")];
            int32 var_1899 = const()[name = string("op_1899"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_37_cast_fp16 = concat(axis = var_1899, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1901_cast_fp16))[name = string("doubled_37_cast_fp16")];
            tensor<int32, [1]> out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_19_gamma_0_to_fp16 = const()[name = string("out_19_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379014336)))];
            fp16 var_1911_to_fp16 = const()[name = string("op_1911_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1911_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1922_split_sizes_0 = const()[name = string("op_1922_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1922_axis_0 = const()[name = string("op_1922_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1922_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1922_cast_fp16_1 = split(axis = var_1922_axis_0, split_sizes = var_1922_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1922_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_1939_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<int32, [2]> var_1945_strides_0 = const()[name = string("op_1945_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1945_pad_type_0 = const()[name = string("op_1945_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1945_pad_0 = const()[name = string("op_1945_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1945_dilations_0 = const()[name = string("op_1945_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1945_groups_0 = const()[name = string("op_1945_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_1945_cast_fp16 = conv(dilations = var_1945_dilations_0, groups = var_1945_groups_0, pad = var_1945_pad_0, pad_type = var_1945_pad_type_0, strides = var_1945_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("op_1945_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_49_cast_fp16 = mul(x = var_1939_cast_fp16, y = var_1945_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_1963_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1963_cast_fp16")];
            int32 var_1961 = const()[name = string("op_1961"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_41_cast_fp16 = concat(axis = var_1961, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1963_cast_fp16))[name = string("doubled_41_cast_fp16")];
            tensor<int32, [1]> out_21_axes_0 = const()[name = string("out_21_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_21_gamma_0_to_fp16 = const()[name = string("out_21_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379022592)))];
            fp16 var_1973_to_fp16 = const()[name = string("op_1973_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1973_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_1984_split_sizes_0 = const()[name = string("op_1984_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1984_axis_0 = const()[name = string("op_1984_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1984_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1984_cast_fp16_1 = split(axis = var_1984_axis_0, split_sizes = var_1984_split_sizes_0, x = out_21_cast_fp16)[name = string("op_1984_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379030848)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_1984_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_2041_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2041_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_2048_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2048_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2052_cast_fp16 = mul(x = x_51_cast_fp16, y = var_308_cast_fp16)[name = string("op_2052_cast_fp16")];
            tensor<int32, [2]> var_2053_split_sizes_0 = const()[name = string("op_2053_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2053_axis_0 = const()[name = string("op_2053_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_2053_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_2053_cast_fp16_1 = split(axis = var_2053_axis_0, split_sizes = var_2053_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2053_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_2055_cast_fp16 = mul(x = var_2053_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2055_cast_fp16")];
            int32 var_2057 = const()[name = string("op_2057"), val = int32(-2)];
            bool var_2058_interleave_0 = const()[name = string("op_2058_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_2058_cast_fp16 = concat(axis = var_2057, interleave = var_2058_interleave_0, values = (var_2055_cast_fp16, var_2053_cast_fp16_0))[name = string("op_2058_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2059_cast_fp16 = mul(x = var_2058_cast_fp16, y = var_315_cast_fp16)[name = string("op_2059_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_33_cast_fp16 = add(x = var_2052_cast_fp16, y = var_2059_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2065_cast_fp16 = mul(x = var_2041_cast_fp16, y = var_308_cast_fp16)[name = string("op_2065_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_2066_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = var_2041_cast_fp16)[name = string("op_2066_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_2068_cast_fp16 = mul(x = var_2066_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2068_cast_fp16")];
            int32 var_2070 = const()[name = string("op_2070"), val = int32(-2)];
            bool var_2071_interleave_0 = const()[name = string("op_2071_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_2071_cast_fp16 = concat(axis = var_2070, interleave = var_2071_interleave_0, values = (var_2068_cast_fp16, var_2066_cast_fp16_0))[name = string("op_2071_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2072_cast_fp16 = mul(x = var_2071_cast_fp16, y = var_315_cast_fp16)[name = string("op_2072_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_55_cast_fp16 = add(x = var_2065_cast_fp16, y = var_2072_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_80)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_82_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_82 = read_state(input = key_cache)[name = string("coreml_update_state_82")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2048_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_81)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_83_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_83 = read_state(input = value_cache)[name = string("coreml_update_state_83")];
            tensor<int32, [4]> var_2142_begin_0 = const()[name = string("op_2142_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2142_end_0 = const()[name = string("op_2142_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2142_end_mask_0 = const()[name = string("op_2142_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2142_cast_fp16 = slice_by_index(begin = var_2142_begin_0, end = var_2142_end_0, end_mask = var_2142_end_mask_0, x = coreml_update_state_82)[name = string("op_2142_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2145_axis_0 = const()[name = string("op_2145_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_1 = split(axis = var_2145_axis_0, split_sizes = tile_10, x = var_2142_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [4]> var_2152_begin_0 = const()[name = string("op_2152_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2152_end_0 = const()[name = string("op_2152_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2152_end_mask_0 = const()[name = string("op_2152_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2152_cast_fp16 = slice_by_index(begin = var_2152_begin_0, end = var_2152_end_0, end_mask = var_2152_end_mask_0, x = coreml_update_state_83)[name = string("op_2152_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2155_axis_0 = const()[name = string("op_2155_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_1 = split(axis = var_2155_axis_0, split_sizes = tile_11, x = var_2152_cast_fp16)[name = string("op_2155_cast_fp16")];
            tensor<int32, [2]> var_2158_split_sizes_0 = const()[name = string("op_2158_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2158_axis_0 = const()[name = string("op_2158_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_2158_0, tensor<fp16, [1, 8, 128, 32]> var_2158_1 = split(axis = var_2158_axis_0, split_sizes = var_2158_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2158")];
            bool attn_weights_81_transpose_x_0 = const()[name = string("attn_weights_81_transpose_x_0"), val = bool(false)];
            bool attn_weights_81_transpose_y_0 = const()[name = string("attn_weights_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_81_cast_fp16 = matmul(transpose_x = attn_weights_81_transpose_x_0, transpose_y = attn_weights_81_transpose_y_0, x = var_2145_cast_fp16_0, y = var_2158_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2161_to_fp16 = const()[name = string("op_2161_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_83_cast_fp16 = mul(x = attn_weights_81_cast_fp16, y = var_2161_to_fp16)[name = string("attn_weights_83_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_85_cast_fp16 = add(x = attn_weights_83_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_85_cast_fp16")];
            int32 var_2165 = const()[name = string("op_2165"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_87_cast_fp16 = softmax(axis = var_2165, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2171_transpose_x_1 = const()[name = string("op_2171_transpose_x_1"), val = bool(true)];
            bool var_2171_transpose_y_1 = const()[name = string("op_2171_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2171_cast_fp16 = matmul(transpose_x = var_2171_transpose_x_1, transpose_y = var_2171_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2155_cast_fp16_0)[name = string("op_2171_cast_fp16")];
            bool attn_weights_89_transpose_x_0 = const()[name = string("attn_weights_89_transpose_x_0"), val = bool(false)];
            bool attn_weights_89_transpose_y_0 = const()[name = string("attn_weights_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_89_cast_fp16 = matmul(transpose_x = attn_weights_89_transpose_x_0, transpose_y = attn_weights_89_transpose_y_0, x = var_2145_cast_fp16_1, y = var_2158_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2173_to_fp16 = const()[name = string("op_2173_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_91_cast_fp16 = mul(x = attn_weights_89_cast_fp16, y = var_2173_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            int32 var_2177 = const()[name = string("op_2177"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_95_cast_fp16 = softmax(axis = var_2177, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")];
            bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(true)];
            bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_95_cast_fp16, y = var_2155_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2185 = const()[name = string("op_2185"), val = int32(1)];
            bool attn_output_43_interleave_0 = const()[name = string("attn_output_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_43_cast_fp16 = concat(axis = var_2185, interleave = attn_output_43_interleave_0, values = (var_2171_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2189_perm_0 = const()[name = string("op_2189_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_2189_cast_fp16 = transpose(perm = var_2189_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2189_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_2222_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2222_cast_fp16")];
            int32 var_2220 = const()[name = string("op_2220"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_45_cast_fp16 = concat(axis = var_2220, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2222_cast_fp16))[name = string("doubled_45_cast_fp16")];
            tensor<int32, [1]> out_23_axes_0 = const()[name = string("out_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_23_gamma_0_to_fp16 = const()[name = string("out_23_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380079488)))];
            fp16 var_2232_to_fp16 = const()[name = string("op_2232_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_2232_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2243_split_sizes_0 = const()[name = string("op_2243_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2243_axis_0 = const()[name = string("op_2243_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2243_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2243_cast_fp16_1 = split(axis = var_2243_axis_0, split_sizes = var_2243_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2243_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_2260_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<int32, [2]> var_2266_strides_0 = const()[name = string("op_2266_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2266_pad_type_0 = const()[name = string("op_2266_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2266_pad_0 = const()[name = string("op_2266_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2266_dilations_0 = const()[name = string("op_2266_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2266_groups_0 = const()[name = string("op_2266_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_2266_cast_fp16 = conv(dilations = var_2266_dilations_0, groups = var_2266_groups_0, pad = var_2266_pad_0, pad_type = var_2266_pad_type_0, strides = var_2266_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("op_2266_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_59_cast_fp16 = mul(x = var_2260_cast_fp16, y = var_2266_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_2284_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2284_cast_fp16")];
            int32 var_2282 = const()[name = string("op_2282"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_49_cast_fp16 = concat(axis = var_2282, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2284_cast_fp16))[name = string("doubled_49_cast_fp16")];
            tensor<int32, [1]> out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_25_gamma_0_to_fp16 = const()[name = string("out_25_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380087744)))];
            fp16 var_2294_to_fp16 = const()[name = string("op_2294_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_2294_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2305_split_sizes_0 = const()[name = string("op_2305_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2305_axis_0 = const()[name = string("op_2305_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2305_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2305_cast_fp16_1 = split(axis = var_2305_axis_0, split_sizes = var_2305_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380096000)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2305_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_2362_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2362_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_2369_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2369_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2373_cast_fp16 = mul(x = x_61_cast_fp16, y = var_308_cast_fp16)[name = string("op_2373_cast_fp16")];
            tensor<int32, [2]> var_2374_split_sizes_0 = const()[name = string("op_2374_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2374_axis_0 = const()[name = string("op_2374_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_2374_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_2374_cast_fp16_1 = split(axis = var_2374_axis_0, split_sizes = var_2374_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2376_cast_fp16")];
            int32 var_2378 = const()[name = string("op_2378"), val = int32(-2)];
            bool var_2379_interleave_0 = const()[name = string("op_2379_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_2379_cast_fp16 = concat(axis = var_2378, interleave = var_2379_interleave_0, values = (var_2376_cast_fp16, var_2374_cast_fp16_0))[name = string("op_2379_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2380_cast_fp16 = mul(x = var_2379_cast_fp16, y = var_315_cast_fp16)[name = string("op_2380_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_39_cast_fp16 = add(x = var_2373_cast_fp16, y = var_2380_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2386_cast_fp16 = mul(x = var_2362_cast_fp16, y = var_308_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_2387_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = var_2362_cast_fp16)[name = string("op_2387_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_2389_cast_fp16 = mul(x = var_2387_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2389_cast_fp16")];
            int32 var_2391 = const()[name = string("op_2391"), val = int32(-2)];
            bool var_2392_interleave_0 = const()[name = string("op_2392_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_2392_cast_fp16 = concat(axis = var_2391, interleave = var_2392_interleave_0, values = (var_2389_cast_fp16, var_2387_cast_fp16_0))[name = string("op_2392_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2393_cast_fp16 = mul(x = var_2392_cast_fp16, y = var_315_cast_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_65_cast_fp16 = add(x = var_2386_cast_fp16, y = var_2393_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_82)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_84_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_84 = read_state(input = key_cache)[name = string("coreml_update_state_84")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2369_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_83)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_85_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_85 = read_state(input = value_cache)[name = string("coreml_update_state_85")];
            tensor<int32, [4]> var_2463_begin_0 = const()[name = string("op_2463_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2463_end_0 = const()[name = string("op_2463_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2463_end_mask_0 = const()[name = string("op_2463_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2463_cast_fp16 = slice_by_index(begin = var_2463_begin_0, end = var_2463_end_0, end_mask = var_2463_end_mask_0, x = coreml_update_state_84)[name = string("op_2463_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2466_axis_0 = const()[name = string("op_2466_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_1 = split(axis = var_2466_axis_0, split_sizes = tile_12, x = var_2463_cast_fp16)[name = string("op_2466_cast_fp16")];
            tensor<int32, [4]> var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, x = coreml_update_state_85)[name = string("op_2473_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2476_axis_0 = const()[name = string("op_2476_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_1 = split(axis = var_2476_axis_0, split_sizes = tile_13, x = var_2473_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<int32, [2]> var_2479_split_sizes_0 = const()[name = string("op_2479_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2479_axis_0 = const()[name = string("op_2479_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_2479_0, tensor<fp16, [1, 8, 128, 32]> var_2479_1 = split(axis = var_2479_axis_0, split_sizes = var_2479_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2479")];
            bool attn_weights_97_transpose_x_0 = const()[name = string("attn_weights_97_transpose_x_0"), val = bool(false)];
            bool attn_weights_97_transpose_y_0 = const()[name = string("attn_weights_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_0, transpose_y = attn_weights_97_transpose_y_0, x = var_2466_cast_fp16_0, y = var_2479_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2482_to_fp16 = const()[name = string("op_2482_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_2482_to_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_101_cast_fp16 = add(x = attn_weights_99_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_101_cast_fp16")];
            int32 var_2486 = const()[name = string("op_2486"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_103_cast_fp16 = softmax(axis = var_2486, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2492_transpose_x_1 = const()[name = string("op_2492_transpose_x_1"), val = bool(true)];
            bool var_2492_transpose_y_1 = const()[name = string("op_2492_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2492_cast_fp16 = matmul(transpose_x = var_2492_transpose_x_1, transpose_y = var_2492_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2476_cast_fp16_0)[name = string("op_2492_cast_fp16")];
            bool attn_weights_105_transpose_x_0 = const()[name = string("attn_weights_105_transpose_x_0"), val = bool(false)];
            bool attn_weights_105_transpose_y_0 = const()[name = string("attn_weights_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_105_cast_fp16 = matmul(transpose_x = attn_weights_105_transpose_x_0, transpose_y = attn_weights_105_transpose_y_0, x = var_2466_cast_fp16_1, y = var_2479_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2494_to_fp16 = const()[name = string("op_2494_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_107_cast_fp16 = mul(x = attn_weights_105_cast_fp16, y = var_2494_to_fp16)[name = string("attn_weights_107_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_109_cast_fp16 = add(x = attn_weights_107_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_109_cast_fp16")];
            int32 var_2498 = const()[name = string("op_2498"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_111_cast_fp16 = softmax(axis = var_2498, x = attn_weights_109_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(true)];
            bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_111_cast_fp16, y = var_2476_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2506 = const()[name = string("op_2506"), val = int32(1)];
            bool attn_output_51_interleave_0 = const()[name = string("attn_output_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_51_cast_fp16 = concat(axis = var_2506, interleave = attn_output_51_interleave_0, values = (var_2492_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2510_perm_0 = const()[name = string("op_2510_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_2510_cast_fp16 = transpose(perm = var_2510_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2510_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_2543_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2543_cast_fp16")];
            int32 var_2541 = const()[name = string("op_2541"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_53_cast_fp16 = concat(axis = var_2541, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2543_cast_fp16))[name = string("doubled_53_cast_fp16")];
            tensor<int32, [1]> out_27_axes_0 = const()[name = string("out_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_27_gamma_0_to_fp16 = const()[name = string("out_27_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381144640)))];
            fp16 var_2553_to_fp16 = const()[name = string("op_2553_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_2553_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2564_split_sizes_0 = const()[name = string("op_2564_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2564_axis_0 = const()[name = string("op_2564_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2564_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2564_cast_fp16_1 = split(axis = var_2564_axis_0, split_sizes = var_2564_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_2581_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [2]> var_2587_strides_0 = const()[name = string("op_2587_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2587_pad_type_0 = const()[name = string("op_2587_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2587_pad_0 = const()[name = string("op_2587_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2587_dilations_0 = const()[name = string("op_2587_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2587_groups_0 = const()[name = string("op_2587_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_2587_cast_fp16 = conv(dilations = var_2587_dilations_0, groups = var_2587_groups_0, pad = var_2587_pad_0, pad_type = var_2587_pad_type_0, strides = var_2587_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("op_2587_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_69_cast_fp16 = mul(x = var_2581_cast_fp16, y = var_2587_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381152896)))];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_to_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_2605_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2605_cast_fp16")];
            int32 var_2603 = const()[name = string("op_2603"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_57_cast_fp16 = concat(axis = var_2603, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2605_cast_fp16))[name = string("doubled_57_cast_fp16")];
            tensor<int32, [1]> out_29_axes_0 = const()[name = string("out_29_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_29_gamma_0_to_fp16 = const()[name = string("out_29_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406318784)))];
            fp16 var_2615_to_fp16 = const()[name = string("op_2615_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_2615_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2626_split_sizes_0 = const()[name = string("op_2626_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2626_axis_0 = const()[name = string("op_2626_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2626_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2626_cast_fp16_1 = split(axis = var_2626_axis_0, split_sizes = var_2626_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2626_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406327040)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2626_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_2683_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_2690_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2690_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2694_cast_fp16 = mul(x = x_71_cast_fp16, y = var_308_cast_fp16)[name = string("op_2694_cast_fp16")];
            tensor<int32, [2]> var_2695_split_sizes_0 = const()[name = string("op_2695_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2695_axis_0 = const()[name = string("op_2695_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_2695_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_2695_cast_fp16_1 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2695_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_2697_cast_fp16 = mul(x = var_2695_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2697_cast_fp16")];
            int32 var_2699 = const()[name = string("op_2699"), val = int32(-2)];
            bool var_2700_interleave_0 = const()[name = string("op_2700_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_2700_cast_fp16 = concat(axis = var_2699, interleave = var_2700_interleave_0, values = (var_2697_cast_fp16, var_2695_cast_fp16_0))[name = string("op_2700_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2701_cast_fp16 = mul(x = var_2700_cast_fp16, y = var_315_cast_fp16)[name = string("op_2701_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_45_cast_fp16 = add(x = var_2694_cast_fp16, y = var_2701_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2707_cast_fp16 = mul(x = var_2683_cast_fp16, y = var_308_cast_fp16)[name = string("op_2707_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_2708_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = var_2683_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_2710_cast_fp16 = mul(x = var_2708_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2710_cast_fp16")];
            int32 var_2712 = const()[name = string("op_2712"), val = int32(-2)];
            bool var_2713_interleave_0 = const()[name = string("op_2713_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_2713_cast_fp16 = concat(axis = var_2712, interleave = var_2713_interleave_0, values = (var_2710_cast_fp16, var_2708_cast_fp16_0))[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2714_cast_fp16 = mul(x = var_2713_cast_fp16, y = var_315_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_75_cast_fp16 = add(x = var_2707_cast_fp16, y = var_2714_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_84)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_86_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_86 = read_state(input = key_cache)[name = string("coreml_update_state_86")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2690_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_85)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_87_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_87 = read_state(input = value_cache)[name = string("coreml_update_state_87")];
            tensor<int32, [4]> var_2784_begin_0 = const()[name = string("op_2784_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2784_end_0 = const()[name = string("op_2784_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2784_end_mask_0 = const()[name = string("op_2784_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2784_cast_fp16 = slice_by_index(begin = var_2784_begin_0, end = var_2784_end_0, end_mask = var_2784_end_mask_0, x = coreml_update_state_86)[name = string("op_2784_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2787_axis_0 = const()[name = string("op_2787_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_1 = split(axis = var_2787_axis_0, split_sizes = tile_14, x = var_2784_cast_fp16)[name = string("op_2787_cast_fp16")];
            tensor<int32, [4]> var_2794_begin_0 = const()[name = string("op_2794_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2794_end_0 = const()[name = string("op_2794_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2794_end_mask_0 = const()[name = string("op_2794_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2794_cast_fp16 = slice_by_index(begin = var_2794_begin_0, end = var_2794_end_0, end_mask = var_2794_end_mask_0, x = coreml_update_state_87)[name = string("op_2794_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2797_axis_0 = const()[name = string("op_2797_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_1 = split(axis = var_2797_axis_0, split_sizes = tile_15, x = var_2794_cast_fp16)[name = string("op_2797_cast_fp16")];
            tensor<int32, [2]> var_2800_split_sizes_0 = const()[name = string("op_2800_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2800_axis_0 = const()[name = string("op_2800_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_2800_0, tensor<fp16, [1, 8, 128, 32]> var_2800_1 = split(axis = var_2800_axis_0, split_sizes = var_2800_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2800")];
            bool attn_weights_113_transpose_x_0 = const()[name = string("attn_weights_113_transpose_x_0"), val = bool(false)];
            bool attn_weights_113_transpose_y_0 = const()[name = string("attn_weights_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_113_cast_fp16 = matmul(transpose_x = attn_weights_113_transpose_x_0, transpose_y = attn_weights_113_transpose_y_0, x = var_2787_cast_fp16_0, y = var_2800_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2803_to_fp16 = const()[name = string("op_2803_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_115_cast_fp16 = mul(x = attn_weights_113_cast_fp16, y = var_2803_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            int32 var_2807 = const()[name = string("op_2807"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_119_cast_fp16 = softmax(axis = var_2807, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2813_transpose_x_1 = const()[name = string("op_2813_transpose_x_1"), val = bool(true)];
            bool var_2813_transpose_y_1 = const()[name = string("op_2813_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2813_cast_fp16 = matmul(transpose_x = var_2813_transpose_x_1, transpose_y = var_2813_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2797_cast_fp16_0)[name = string("op_2813_cast_fp16")];
            bool attn_weights_121_transpose_x_0 = const()[name = string("attn_weights_121_transpose_x_0"), val = bool(false)];
            bool attn_weights_121_transpose_y_0 = const()[name = string("attn_weights_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_0, transpose_y = attn_weights_121_transpose_y_0, x = var_2787_cast_fp16_1, y = var_2800_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2815_to_fp16 = const()[name = string("op_2815_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_2815_to_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_125_cast_fp16 = add(x = attn_weights_123_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_125_cast_fp16")];
            int32 var_2819 = const()[name = string("op_2819"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_127_cast_fp16 = softmax(axis = var_2819, x = attn_weights_125_cast_fp16)[name = string("attn_weights_127_cast_fp16")];
            bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(true)];
            bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_127_cast_fp16, y = var_2797_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2827 = const()[name = string("op_2827"), val = int32(1)];
            bool attn_output_59_interleave_0 = const()[name = string("attn_output_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_59_cast_fp16 = concat(axis = var_2827, interleave = attn_output_59_interleave_0, values = (var_2813_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2831_perm_0 = const()[name = string("op_2831_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_2831_cast_fp16 = transpose(perm = var_2831_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2831_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_2864_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2864_cast_fp16")];
            int32 var_2862 = const()[name = string("op_2862"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_61_cast_fp16 = concat(axis = var_2862, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2864_cast_fp16))[name = string("doubled_61_cast_fp16")];
            tensor<int32, [1]> out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_31_gamma_0_to_fp16 = const()[name = string("out_31_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407375680)))];
            fp16 var_2874_to_fp16 = const()[name = string("op_2874_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_2874_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2885_split_sizes_0 = const()[name = string("op_2885_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2885_axis_0 = const()[name = string("op_2885_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2885_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2885_cast_fp16_1 = split(axis = var_2885_axis_0, split_sizes = var_2885_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2885_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2885_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_2902_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2902_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407383936)))];
            tensor<int32, [2]> var_2908_strides_0 = const()[name = string("op_2908_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2908_pad_type_0 = const()[name = string("op_2908_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2908_pad_0 = const()[name = string("op_2908_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2908_dilations_0 = const()[name = string("op_2908_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2908_groups_0 = const()[name = string("op_2908_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_2908_cast_fp16 = conv(dilations = var_2908_dilations_0, groups = var_2908_groups_0, pad = var_2908_pad_0, pad_type = var_2908_pad_type_0, strides = var_2908_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2885_cast_fp16_0)[name = string("op_2908_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_79_cast_fp16 = mul(x = var_2902_cast_fp16, y = var_2908_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432549824)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_2926_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2926_cast_fp16")];
            int32 var_2924 = const()[name = string("op_2924"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_65_cast_fp16 = concat(axis = var_2924, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2926_cast_fp16))[name = string("doubled_65_cast_fp16")];
            tensor<int32, [1]> out_33_axes_0 = const()[name = string("out_33_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_33_gamma_0_to_fp16 = const()[name = string("out_33_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457715712)))];
            fp16 var_2936_to_fp16 = const()[name = string("op_2936_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_2936_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2947_split_sizes_0 = const()[name = string("op_2947_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2947_axis_0 = const()[name = string("op_2947_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2947_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2947_cast_fp16_1 = split(axis = var_2947_axis_0, split_sizes = var_2947_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2947_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457723968)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466112640)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467161280)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_3004_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3004_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_3011_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3011_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3015_cast_fp16 = mul(x = x_81_cast_fp16, y = var_308_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<int32, [2]> var_3016_split_sizes_0 = const()[name = string("op_3016_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3016_axis_0 = const()[name = string("op_3016_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_3016_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_3016_cast_fp16_1 = split(axis = var_3016_axis_0, split_sizes = var_3016_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3016_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_3018_cast_fp16 = mul(x = var_3016_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3018_cast_fp16")];
            int32 var_3020 = const()[name = string("op_3020"), val = int32(-2)];
            bool var_3021_interleave_0 = const()[name = string("op_3021_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_3021_cast_fp16 = concat(axis = var_3020, interleave = var_3021_interleave_0, values = (var_3018_cast_fp16, var_3016_cast_fp16_0))[name = string("op_3021_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3022_cast_fp16 = mul(x = var_3021_cast_fp16, y = var_315_cast_fp16)[name = string("op_3022_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_51_cast_fp16 = add(x = var_3015_cast_fp16, y = var_3022_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3028_cast_fp16 = mul(x = var_3004_cast_fp16, y = var_308_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_3029_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = var_3004_cast_fp16)[name = string("op_3029_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_3031_cast_fp16 = mul(x = var_3029_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3031_cast_fp16")];
            int32 var_3033 = const()[name = string("op_3033"), val = int32(-2)];
            bool var_3034_interleave_0 = const()[name = string("op_3034_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_3034_cast_fp16 = concat(axis = var_3033, interleave = var_3034_interleave_0, values = (var_3031_cast_fp16, var_3029_cast_fp16_0))[name = string("op_3034_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3035_cast_fp16 = mul(x = var_3034_cast_fp16, y = var_315_cast_fp16)[name = string("op_3035_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_85_cast_fp16 = add(x = var_3028_cast_fp16, y = var_3035_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_86)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_88_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_88 = read_state(input = key_cache)[name = string("coreml_update_state_88")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3011_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_87)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_89_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_89 = read_state(input = value_cache)[name = string("coreml_update_state_89")];
            tensor<int32, [4]> var_3105_begin_0 = const()[name = string("op_3105_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3105_end_0 = const()[name = string("op_3105_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3105_end_mask_0 = const()[name = string("op_3105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3105_cast_fp16 = slice_by_index(begin = var_3105_begin_0, end = var_3105_end_0, end_mask = var_3105_end_mask_0, x = coreml_update_state_88)[name = string("op_3105_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3108_axis_0 = const()[name = string("op_3108_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_1 = split(axis = var_3108_axis_0, split_sizes = tile_16, x = var_3105_cast_fp16)[name = string("op_3108_cast_fp16")];
            tensor<int32, [4]> var_3115_begin_0 = const()[name = string("op_3115_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3115_end_0 = const()[name = string("op_3115_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3115_end_mask_0 = const()[name = string("op_3115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3115_cast_fp16 = slice_by_index(begin = var_3115_begin_0, end = var_3115_end_0, end_mask = var_3115_end_mask_0, x = coreml_update_state_89)[name = string("op_3115_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3118_axis_0 = const()[name = string("op_3118_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_1 = split(axis = var_3118_axis_0, split_sizes = tile_17, x = var_3115_cast_fp16)[name = string("op_3118_cast_fp16")];
            tensor<int32, [2]> var_3121_split_sizes_0 = const()[name = string("op_3121_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3121_axis_0 = const()[name = string("op_3121_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_3121_0, tensor<fp16, [1, 8, 128, 32]> var_3121_1 = split(axis = var_3121_axis_0, split_sizes = var_3121_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3121")];
            bool attn_weights_129_transpose_x_0 = const()[name = string("attn_weights_129_transpose_x_0"), val = bool(false)];
            bool attn_weights_129_transpose_y_0 = const()[name = string("attn_weights_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_129_cast_fp16 = matmul(transpose_x = attn_weights_129_transpose_x_0, transpose_y = attn_weights_129_transpose_y_0, x = var_3108_cast_fp16_0, y = var_3121_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3124_to_fp16 = const()[name = string("op_3124_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_131_cast_fp16 = mul(x = attn_weights_129_cast_fp16, y = var_3124_to_fp16)[name = string("attn_weights_131_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_133_cast_fp16 = add(x = attn_weights_131_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_133_cast_fp16")];
            int32 var_3128 = const()[name = string("op_3128"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_135_cast_fp16 = softmax(axis = var_3128, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3134_transpose_x_1 = const()[name = string("op_3134_transpose_x_1"), val = bool(true)];
            bool var_3134_transpose_y_1 = const()[name = string("op_3134_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3134_cast_fp16 = matmul(transpose_x = var_3134_transpose_x_1, transpose_y = var_3134_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3118_cast_fp16_0)[name = string("op_3134_cast_fp16")];
            bool attn_weights_137_transpose_x_0 = const()[name = string("attn_weights_137_transpose_x_0"), val = bool(false)];
            bool attn_weights_137_transpose_y_0 = const()[name = string("attn_weights_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_137_cast_fp16 = matmul(transpose_x = attn_weights_137_transpose_x_0, transpose_y = attn_weights_137_transpose_y_0, x = var_3108_cast_fp16_1, y = var_3121_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3136_to_fp16 = const()[name = string("op_3136_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_139_cast_fp16 = mul(x = attn_weights_137_cast_fp16, y = var_3136_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            int32 var_3140 = const()[name = string("op_3140"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_cast_fp16 = softmax(axis = var_3140, x = attn_weights_141_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(true)];
            bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_cast_fp16, y = var_3118_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3148 = const()[name = string("op_3148"), val = int32(1)];
            bool attn_output_67_interleave_0 = const()[name = string("attn_output_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_67_cast_fp16 = concat(axis = var_3148, interleave = attn_output_67_interleave_0, values = (var_3134_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3152_perm_0 = const()[name = string("op_3152_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_3152_cast_fp16 = transpose(perm = var_3152_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_cast_fp16 = reshape(shape = concat_107x, x = var_3152_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468209920)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_3185_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3185_cast_fp16")];
            int32 var_3183 = const()[name = string("op_3183"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_69_cast_fp16 = concat(axis = var_3183, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3185_cast_fp16))[name = string("doubled_69_cast_fp16")];
            tensor<int32, [1]> out_35_axes_0 = const()[name = string("out_35_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_35_gamma_0_to_fp16 = const()[name = string("out_35_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476598592)))];
            fp16 var_3195_to_fp16 = const()[name = string("op_3195_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_3195_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3206_split_sizes_0 = const()[name = string("op_3206_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3206_axis_0 = const()[name = string("op_3206_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3206_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3206_cast_fp16_1 = split(axis = var_3206_axis_0, split_sizes = var_3206_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3206_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476606848)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_8_mlp_gate_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_3223_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3223_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501772736)))];
            tensor<int32, [2]> var_3229_strides_0 = const()[name = string("op_3229_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3229_pad_type_0 = const()[name = string("op_3229_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3229_pad_0 = const()[name = string("op_3229_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3229_dilations_0 = const()[name = string("op_3229_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3229_groups_0 = const()[name = string("op_3229_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_3229_cast_fp16 = conv(dilations = var_3229_dilations_0, groups = var_3229_groups_0, pad = var_3229_pad_0, pad_type = var_3229_pad_type_0, strides = var_3229_strides_0, weight = layers_8_mlp_up_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("op_3229_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_cast_fp16 = mul(x = var_3223_cast_fp16, y = var_3229_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526938624)))];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_3247_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3247_cast_fp16")];
            int32 var_3245 = const()[name = string("op_3245"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_73_cast_fp16 = concat(axis = var_3245, interleave = doubled_73_interleave_0, values = (hidden_states_cast_fp16, var_3247_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_axes_0 = const()[name = string("out_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_gamma_0_to_fp16 = const()[name = string("out_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552104512)))];
            fp16 var_3257_to_fp16 = const()[name = string("op_3257_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3257_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3268_split_sizes_0 = const()[name = string("op_3268_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3268_axis_0 = const()[name = string("op_3268_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states, tensor<fp16, [1, 2048, 1, 32]> var_3268_cast_fp16_1 = split(axis = var_3268_axis_0, split_sizes = var_3268_split_sizes_0, x = out_cast_fp16)[name = string("op_3268_cast_fp16")];
        } -> (hidden_states);
    func length_64<ios18>(tensor<fp32, [1, 2048, 1, 64]> inputs_embeds, state<tensor<fp16, [9, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [64]> position_index_seed, state<tensor<fp16, [9, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725952))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17321280))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17308928))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17327488))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29922816))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29910464))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29929024))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42516160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42512000))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42518272))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243840))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243264))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47244160))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51442688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51438528))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444800))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64040128))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64027776))))[name = string("layers_1_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64046336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76633472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76629312))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76635584))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80834112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80829952))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80836224))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361152))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81360576))))[name = string("layers_2_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361472))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85560000))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85555840))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85562112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98145088))))[name = string("layers_2_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110758976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110746624))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110765184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123352320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123348160))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123354432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127552960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127548800))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127555072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128079424))))[name = string("layers_3_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080320))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140675648))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140663296))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140681856))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153277184))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153264832))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153283392))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165870528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165866368))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165872640))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170071168))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170067008))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170073280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170597632))))[name = string("layers_4_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174797056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174792896))))[name = string("layers_4_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174799168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187394496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187382144))))[name = string("layers_4_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187400704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199996032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199983680))))[name = string("layers_4_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_4_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200002240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212589376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212585216))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212591488))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216790016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216785856))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216792128))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317056))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217316480))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317376))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221515904))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221511744))))[name = string("layers_5_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221518016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234113344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234100992))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234119552))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246714880))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246702528))))[name = string("layers_5_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_5_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246721088))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259308224))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259304064))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259310336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263508864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263504704))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263510976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035328))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264036224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268234752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268230592))))[name = string("layers_6_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268236864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280832192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280819840))))[name = string("layers_6_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280838400))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293433728))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293421376))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293439936))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297638464))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297634304))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297640576))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165504))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298164928))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165824))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302364352))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302360192))))[name = string("layers_7_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302366464))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314961792))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314949440))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(64)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [64]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_280 = const()[name = string("op_280"), val = int32(0)];
            bool var_282_exclusive_0 = const()[name = string("op_282_exclusive_0"), val = bool(false)];
            bool var_282_reverse_0 = const()[name = string("op_282_reverse_0"), val = bool(false)];
            tensor<fp16, [64]> var_282_cast_fp16 = cumsum(axis = var_280, exclusive = var_282_exclusive_0, reverse = var_282_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_282_cast_fp16")];
            fp16 var_284_promoted_to_fp16 = const()[name = string("op_284_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [64]> position_offsets_cast_fp16 = sub(x = var_282_cast_fp16, y = var_284_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 64]> var_287_cast_fp16 = expand_dims(axes = var_287_axes_0, x = position_offsets_cast_fp16)[name = string("op_287_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_23")];
            tensor<fp16, [1, 64]> position_ids_1_cast_fp16 = add(x = var_287_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 64]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_22")];
            tensor<bool, [1, 64]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 64]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 64]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314968000)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_10_y_0 = const()[name = string("greater_equal_10_y_0"), val = int32(0)];
            tensor<bool, [1, 64]> greater_equal_10 = greater_equal(x = select_0, y = greater_equal_10_y_0)[name = string("greater_equal_10")];
            int32 slice_by_index_10 = const()[name = string("slice_by_index_10"), val = int32(32768)];
            tensor<int32, [1, 64]> add_10 = add(x = select_0, y = slice_by_index_10)[name = string("add_10")];
            tensor<int32, [1, 64]> select_10 = select(a = select_0, b = add_10, cond = greater_equal_10)[name = string("select_10")];
            int32 cos_1_cast_fp16_axis_5 = const()[name = string("cos_1_cast_fp16_axis_5"), val = int32(0)];
            tensor<fp16, [1, 64, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_5, batch_dims = cos_1_batch_dims_0, indices = select_10, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323356672)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_5 = const()[name = string("sin_1_cast_fp16_axis_5"), val = int32(0)];
            tensor<fp16, [1, 64, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_5, batch_dims = sin_1_batch_dims_0, indices = select_10, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_306_perm_0 = const()[name = string("op_306_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_308_axes_0 = const()[name = string("op_308_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 64]> var_306_cast_fp16 = transpose(perm = var_306_perm_0, x = cos_1_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 1, 128, 64]> var_308_cast_fp16 = expand_dims(axes = var_308_axes_0, x = var_306_cast_fp16)[name = string("op_308_cast_fp16")];
            tensor<int32, [3]> var_313_perm_0 = const()[name = string("op_313_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 64]> var_313_cast_fp16 = transpose(perm = var_313_perm_0, x = sin_1_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [1, 1, 128, 64]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = var_313_cast_fp16)[name = string("op_315_cast_fp16")];
            tensor<int32, [1]> var_334_axes_0 = const()[name = string("op_334_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 64, 1]> var_334 = expand_dims(axes = var_334_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_334")];
            tensor<int32, [1, 1, 2048]> var_327 = const()[name = string("op_327"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331745344)))];
            tensor<bool, [1, 64, 2048]> var_335 = greater(x = var_327, y = var_334)[name = string("op_335")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_342_axes_0 = const()[name = string("op_342_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 64, 2048]> var_335_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_335)[name = string("cast_21")];
            tensor<fp16, [1, 1, 64, 2048]> var_342_cast_fp16 = expand_dims(axes = var_342_axes_0, x = var_335_to_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_346_promoted_to_fp16 = const()[name = string("op_346_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 64]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_342_cast_fp16)[name = string("transpose_177")];
            tensor<bool, [1, 1, 2048, 64]> var_347_cast_fp16 = equal(x = mask_cast_fp16, y = var_346_promoted_to_fp16)[name = string("op_347_cast_fp16")];
            fp16 var_348_to_fp16 = const()[name = string("op_348_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 64]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_348_to_fp16, cond = var_347_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_20")];
            tensor<fp16, [1, 2048, 1, 64]> var_358_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_358_cast_fp16")];
            int32 var_356 = const()[name = string("op_356"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_1_cast_fp16 = concat(axis = var_356, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_358_cast_fp16))[name = string("doubled_1_cast_fp16")];
            tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_1_gamma_0_to_fp16 = const()[name = string("out_1_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331753600)))];
            fp16 var_368_to_fp16 = const()[name = string("op_368_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_368_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_379_split_sizes_0 = const()[name = string("op_379_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_379_axis_0 = const()[name = string("op_379_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_379_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_379_cast_fp16_1 = split(axis = var_379_axis_0, split_sizes = var_379_split_sizes_0, x = out_1_cast_fp16)[name = string("op_379_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331761856)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_379_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_436_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_436_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_443_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_443_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_447_cast_fp16 = mul(x = x_1_cast_fp16, y = var_308_cast_fp16)[name = string("op_447_cast_fp16")];
            tensor<int32, [2]> var_448_split_sizes_0 = const()[name = string("op_448_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_448_axis_0 = const()[name = string("op_448_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_448_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_448_cast_fp16_1 = split(axis = var_448_axis_0, split_sizes = var_448_split_sizes_0, x = x_1_cast_fp16)[name = string("op_448_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_450_cast_fp16 = mul(x = var_448_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_450_cast_fp16")];
            int32 var_452 = const()[name = string("op_452"), val = int32(-2)];
            bool var_453_interleave_0 = const()[name = string("op_453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_453_cast_fp16 = concat(axis = var_452, interleave = var_453_interleave_0, values = (var_450_cast_fp16, var_448_cast_fp16_0))[name = string("op_453_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_454_cast_fp16 = mul(x = var_453_cast_fp16, y = var_315_cast_fp16)[name = string("op_454_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_3_cast_fp16 = add(x = var_447_cast_fp16, y = var_454_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_460_cast_fp16 = mul(x = var_436_cast_fp16, y = var_308_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [2]> var_461_split_sizes_0 = const()[name = string("op_461_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_461_axis_0 = const()[name = string("op_461_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_461_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_461_cast_fp16_1 = split(axis = var_461_axis_0, split_sizes = var_461_split_sizes_0, x = var_436_cast_fp16)[name = string("op_461_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_463_cast_fp16 = mul(x = var_461_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_463_cast_fp16")];
            int32 var_465 = const()[name = string("op_465"), val = int32(-2)];
            bool var_466_interleave_0 = const()[name = string("op_466_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_466_cast_fp16 = concat(axis = var_465, interleave = var_466_interleave_0, values = (var_463_cast_fp16, var_461_cast_fp16_0))[name = string("op_466_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_467_cast_fp16 = mul(x = var_466_cast_fp16, y = var_315_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_5_cast_fp16 = add(x = var_460_cast_fp16, y = var_467_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_176")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_90_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_90 = read_state(input = key_cache)[name = string("coreml_update_state_90")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_443_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_91_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_91 = read_state(input = value_cache)[name = string("coreml_update_state_91")];
            tensor<int32, [4]> var_537_begin_0 = const()[name = string("op_537_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_537_end_0 = const()[name = string("op_537_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_537_end_mask_0 = const()[name = string("op_537_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_537_cast_fp16 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = coreml_update_state_90)[name = string("op_537_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_540_axis_0 = const()[name = string("op_540_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_1 = split(axis = var_540_axis_0, split_sizes = tile_0, x = var_537_cast_fp16)[name = string("op_540_cast_fp16")];
            tensor<int32, [4]> var_547_begin_0 = const()[name = string("op_547_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_547_end_0 = const()[name = string("op_547_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_547_end_mask_0 = const()[name = string("op_547_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_547_cast_fp16 = slice_by_index(begin = var_547_begin_0, end = var_547_end_0, end_mask = var_547_end_mask_0, x = coreml_update_state_91)[name = string("op_547_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_550_axis_0 = const()[name = string("op_550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_1 = split(axis = var_550_axis_0, split_sizes = tile_1, x = var_547_cast_fp16)[name = string("op_550_cast_fp16")];
            tensor<int32, [2]> var_553_split_sizes_0 = const()[name = string("op_553_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_553_axis_0 = const()[name = string("op_553_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_553_0, tensor<fp16, [1, 8, 128, 64]> var_553_1 = split(axis = var_553_axis_0, split_sizes = var_553_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_553")];
            bool attn_weights_1_transpose_x_0 = const()[name = string("attn_weights_1_transpose_x_0"), val = bool(false)];
            bool attn_weights_1_transpose_y_0 = const()[name = string("attn_weights_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_540_cast_fp16_0, y = var_553_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_556_to_fp16 = const()[name = string("op_556_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_556_to_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_5_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
            int32 var_560 = const()[name = string("op_560"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_7_cast_fp16 = softmax(axis = var_560, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_566_transpose_x_1 = const()[name = string("op_566_transpose_x_1"), val = bool(true)];
            bool var_566_transpose_y_1 = const()[name = string("op_566_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_566_cast_fp16 = matmul(transpose_x = var_566_transpose_x_1, transpose_y = var_566_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_550_cast_fp16_0)[name = string("op_566_cast_fp16")];
            bool attn_weights_9_transpose_x_0 = const()[name = string("attn_weights_9_transpose_x_0"), val = bool(false)];
            bool attn_weights_9_transpose_y_0 = const()[name = string("attn_weights_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_9_cast_fp16 = matmul(transpose_x = attn_weights_9_transpose_x_0, transpose_y = attn_weights_9_transpose_y_0, x = var_540_cast_fp16_1, y = var_553_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_568_to_fp16 = const()[name = string("op_568_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_11_cast_fp16 = mul(x = attn_weights_9_cast_fp16, y = var_568_to_fp16)[name = string("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_13_cast_fp16 = add(x = attn_weights_11_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_13_cast_fp16")];
            int32 var_572 = const()[name = string("op_572"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_15_cast_fp16 = softmax(axis = var_572, x = attn_weights_13_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(true)];
            bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_15_cast_fp16, y = var_550_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_580 = const()[name = string("op_580"), val = int32(1)];
            bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_3_cast_fp16 = concat(axis = var_580, interleave = attn_output_3_interleave_0, values = (var_566_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_584_perm_0 = const()[name = string("op_584_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_584_cast_fp16 = transpose(perm = var_584_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_584_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332810496)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_617_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_617_cast_fp16")];
            int32 var_615 = const()[name = string("op_615"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_5_cast_fp16 = concat(axis = var_615, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_617_cast_fp16))[name = string("doubled_5_cast_fp16")];
            tensor<int32, [1]> out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_3_gamma_0_to_fp16 = const()[name = string("out_3_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341199168)))];
            fp16 var_627_to_fp16 = const()[name = string("op_627_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_627_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_638_split_sizes_0 = const()[name = string("op_638_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_638_axis_0 = const()[name = string("op_638_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_638_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_638_cast_fp16_1 = split(axis = var_638_axis_0, split_sizes = var_638_split_sizes_0, x = out_3_cast_fp16)[name = string("op_638_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_655_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_655_cast_fp16")];
            tensor<int32, [2]> var_661_strides_0 = const()[name = string("op_661_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_661_pad_type_0 = const()[name = string("op_661_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_661_pad_0 = const()[name = string("op_661_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_661_dilations_0 = const()[name = string("op_661_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_661_groups_0 = const()[name = string("op_661_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_661_cast_fp16 = conv(dilations = var_661_dilations_0, groups = var_661_groups_0, pad = var_661_pad_0, pad_type = var_661_pad_type_0, strides = var_661_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("op_661_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_9_cast_fp16 = mul(x = var_655_cast_fp16, y = var_661_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_679_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_679_cast_fp16")];
            int32 var_677 = const()[name = string("op_677"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_9_cast_fp16 = concat(axis = var_677, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_679_cast_fp16))[name = string("doubled_9_cast_fp16")];
            tensor<int32, [1]> out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_5_gamma_0_to_fp16 = const()[name = string("out_5_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341207424)))];
            fp16 var_689_to_fp16 = const()[name = string("op_689_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_689_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_700_split_sizes_0 = const()[name = string("op_700_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_700_axis_0 = const()[name = string("op_700_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_700_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_700_cast_fp16_1 = split(axis = var_700_axis_0, split_sizes = var_700_split_sizes_0, x = out_5_cast_fp16)[name = string("op_700_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341215680)))];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = var_700_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_757_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_757_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_764_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_764_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_768_cast_fp16 = mul(x = x_11_cast_fp16, y = var_308_cast_fp16)[name = string("op_768_cast_fp16")];
            tensor<int32, [2]> var_769_split_sizes_0 = const()[name = string("op_769_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_769_axis_0 = const()[name = string("op_769_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_769_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_769_cast_fp16_1 = split(axis = var_769_axis_0, split_sizes = var_769_split_sizes_0, x = x_11_cast_fp16)[name = string("op_769_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_771_cast_fp16 = mul(x = var_769_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_771_cast_fp16")];
            int32 var_773 = const()[name = string("op_773"), val = int32(-2)];
            bool var_774_interleave_0 = const()[name = string("op_774_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_774_cast_fp16 = concat(axis = var_773, interleave = var_774_interleave_0, values = (var_771_cast_fp16, var_769_cast_fp16_0))[name = string("op_774_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_775_cast_fp16 = mul(x = var_774_cast_fp16, y = var_315_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_9_cast_fp16 = add(x = var_768_cast_fp16, y = var_775_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_781_cast_fp16 = mul(x = var_757_cast_fp16, y = var_308_cast_fp16)[name = string("op_781_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_782_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = var_757_cast_fp16)[name = string("op_782_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_784_cast_fp16 = mul(x = var_782_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_784_cast_fp16")];
            int32 var_786 = const()[name = string("op_786"), val = int32(-2)];
            bool var_787_interleave_0 = const()[name = string("op_787_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_787_cast_fp16 = concat(axis = var_786, interleave = var_787_interleave_0, values = (var_784_cast_fp16, var_782_cast_fp16_0))[name = string("op_787_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_788_cast_fp16 = mul(x = var_787_cast_fp16, y = var_315_cast_fp16)[name = string("op_788_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_15_cast_fp16 = add(x = var_781_cast_fp16, y = var_788_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_173")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_90)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_92_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_92 = read_state(input = key_cache)[name = string("coreml_update_state_92")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_764_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_91)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_93_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_93 = read_state(input = value_cache)[name = string("coreml_update_state_93")];
            tensor<int32, [4]> var_858_begin_0 = const()[name = string("op_858_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_858_end_0 = const()[name = string("op_858_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_858_end_mask_0 = const()[name = string("op_858_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_858_cast_fp16 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = coreml_update_state_92)[name = string("op_858_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_861_axis_0 = const()[name = string("op_861_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_1 = split(axis = var_861_axis_0, split_sizes = tile_2, x = var_858_cast_fp16)[name = string("op_861_cast_fp16")];
            tensor<int32, [4]> var_868_begin_0 = const()[name = string("op_868_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_868_end_0 = const()[name = string("op_868_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_868_end_mask_0 = const()[name = string("op_868_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_868_cast_fp16 = slice_by_index(begin = var_868_begin_0, end = var_868_end_0, end_mask = var_868_end_mask_0, x = coreml_update_state_93)[name = string("op_868_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_871_axis_0 = const()[name = string("op_871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_1 = split(axis = var_871_axis_0, split_sizes = tile_3, x = var_868_cast_fp16)[name = string("op_871_cast_fp16")];
            tensor<int32, [2]> var_874_split_sizes_0 = const()[name = string("op_874_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_874_axis_0 = const()[name = string("op_874_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_874_0, tensor<fp16, [1, 8, 128, 64]> var_874_1 = split(axis = var_874_axis_0, split_sizes = var_874_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_874")];
            bool attn_weights_17_transpose_x_0 = const()[name = string("attn_weights_17_transpose_x_0"), val = bool(false)];
            bool attn_weights_17_transpose_y_0 = const()[name = string("attn_weights_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_17_cast_fp16 = matmul(transpose_x = attn_weights_17_transpose_x_0, transpose_y = attn_weights_17_transpose_y_0, x = var_861_cast_fp16_0, y = var_874_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_877_to_fp16 = const()[name = string("op_877_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_19_cast_fp16 = mul(x = attn_weights_17_cast_fp16, y = var_877_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            int32 var_881 = const()[name = string("op_881"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_23_cast_fp16 = softmax(axis = var_881, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_887_transpose_x_1 = const()[name = string("op_887_transpose_x_1"), val = bool(true)];
            bool var_887_transpose_y_1 = const()[name = string("op_887_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_887_cast_fp16 = matmul(transpose_x = var_887_transpose_x_1, transpose_y = var_887_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_871_cast_fp16_0)[name = string("op_887_cast_fp16")];
            bool attn_weights_25_transpose_x_0 = const()[name = string("attn_weights_25_transpose_x_0"), val = bool(false)];
            bool attn_weights_25_transpose_y_0 = const()[name = string("attn_weights_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_0, transpose_y = attn_weights_25_transpose_y_0, x = var_861_cast_fp16_1, y = var_874_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_889_to_fp16 = const()[name = string("op_889_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_889_to_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_29_cast_fp16 = add(x = attn_weights_27_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
            int32 var_893 = const()[name = string("op_893"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_31_cast_fp16 = softmax(axis = var_893, x = attn_weights_29_cast_fp16)[name = string("attn_weights_31_cast_fp16")];
            bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(true)];
            bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_31_cast_fp16, y = var_871_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_901 = const()[name = string("op_901"), val = int32(1)];
            bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_11_cast_fp16 = concat(axis = var_901, interleave = attn_output_11_interleave_0, values = (var_887_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_905_perm_0 = const()[name = string("op_905_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_905_cast_fp16 = transpose(perm = var_905_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_905_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_938_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_938_cast_fp16")];
            int32 var_936 = const()[name = string("op_936"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_13_cast_fp16 = concat(axis = var_936, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_938_cast_fp16))[name = string("doubled_13_cast_fp16")];
            tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_7_gamma_0_to_fp16 = const()[name = string("out_7_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342264320)))];
            fp16 var_948_to_fp16 = const()[name = string("op_948_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_948_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_959_split_sizes_0 = const()[name = string("op_959_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_959_axis_0 = const()[name = string("op_959_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_959_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_959_cast_fp16_1 = split(axis = var_959_axis_0, split_sizes = var_959_split_sizes_0, x = out_7_cast_fp16)[name = string("op_959_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_959_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_976_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_976_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342272576)))];
            tensor<int32, [2]> var_982_strides_0 = const()[name = string("op_982_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_982_pad_type_0 = const()[name = string("op_982_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_982_pad_0 = const()[name = string("op_982_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_982_dilations_0 = const()[name = string("op_982_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_982_groups_0 = const()[name = string("op_982_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_982_cast_fp16 = conv(dilations = var_982_dilations_0, groups = var_982_groups_0, pad = var_982_pad_0, pad_type = var_982_pad_type_0, strides = var_982_strides_0, weight = layers_1_mlp_up_proj_weight_to_fp16, x = var_959_cast_fp16_0)[name = string("op_982_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_19_cast_fp16 = mul(x = var_976_cast_fp16, y = var_982_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1000_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1000_cast_fp16")];
            int32 var_998 = const()[name = string("op_998"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_17_cast_fp16 = concat(axis = var_998, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1000_cast_fp16))[name = string("doubled_17_cast_fp16")];
            tensor<int32, [1]> out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_9_gamma_0_to_fp16 = const()[name = string("out_9_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367438464)))];
            fp16 var_1010_to_fp16 = const()[name = string("op_1010_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1010_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1021_split_sizes_0 = const()[name = string("op_1021_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1021_axis_0 = const()[name = string("op_1021_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1021_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1021_cast_fp16_1 = split(axis = var_1021_axis_0, split_sizes = var_1021_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1021_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367446720)))];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = var_1021_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_1078_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1078_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_1085_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1085_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1089_cast_fp16 = mul(x = x_21_cast_fp16, y = var_308_cast_fp16)[name = string("op_1089_cast_fp16")];
            tensor<int32, [2]> var_1090_split_sizes_0 = const()[name = string("op_1090_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1090_axis_0 = const()[name = string("op_1090_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_1090_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1090_cast_fp16_1 = split(axis = var_1090_axis_0, split_sizes = var_1090_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1090_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_1092_cast_fp16 = mul(x = var_1090_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            int32 var_1094 = const()[name = string("op_1094"), val = int32(-2)];
            bool var_1095_interleave_0 = const()[name = string("op_1095_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1095_cast_fp16 = concat(axis = var_1094, interleave = var_1095_interleave_0, values = (var_1092_cast_fp16, var_1090_cast_fp16_0))[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1096_cast_fp16 = mul(x = var_1095_cast_fp16, y = var_315_cast_fp16)[name = string("op_1096_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_15_cast_fp16 = add(x = var_1089_cast_fp16, y = var_1096_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1102_cast_fp16 = mul(x = var_1078_cast_fp16, y = var_308_cast_fp16)[name = string("op_1102_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_1103_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = var_1078_cast_fp16)[name = string("op_1103_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_1105_cast_fp16 = mul(x = var_1103_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1105_cast_fp16")];
            int32 var_1107 = const()[name = string("op_1107"), val = int32(-2)];
            bool var_1108_interleave_0 = const()[name = string("op_1108_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1108_cast_fp16 = concat(axis = var_1107, interleave = var_1108_interleave_0, values = (var_1105_cast_fp16, var_1103_cast_fp16_0))[name = string("op_1108_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1109_cast_fp16 = mul(x = var_1108_cast_fp16, y = var_315_cast_fp16)[name = string("op_1109_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_25_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1109_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_92)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_94_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_94 = read_state(input = key_cache)[name = string("coreml_update_state_94")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1085_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_93)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_95_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_95 = read_state(input = value_cache)[name = string("coreml_update_state_95")];
            tensor<int32, [4]> var_1179_begin_0 = const()[name = string("op_1179_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1179_end_0 = const()[name = string("op_1179_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1179_end_mask_0 = const()[name = string("op_1179_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1179_cast_fp16 = slice_by_index(begin = var_1179_begin_0, end = var_1179_end_0, end_mask = var_1179_end_mask_0, x = coreml_update_state_94)[name = string("op_1179_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1182_axis_0 = const()[name = string("op_1182_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_1 = split(axis = var_1182_axis_0, split_sizes = tile_4, x = var_1179_cast_fp16)[name = string("op_1182_cast_fp16")];
            tensor<int32, [4]> var_1189_begin_0 = const()[name = string("op_1189_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1189_end_0 = const()[name = string("op_1189_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1189_end_mask_0 = const()[name = string("op_1189_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1189_cast_fp16 = slice_by_index(begin = var_1189_begin_0, end = var_1189_end_0, end_mask = var_1189_end_mask_0, x = coreml_update_state_95)[name = string("op_1189_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1192_axis_0 = const()[name = string("op_1192_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_1 = split(axis = var_1192_axis_0, split_sizes = tile_5, x = var_1189_cast_fp16)[name = string("op_1192_cast_fp16")];
            tensor<int32, [2]> var_1195_split_sizes_0 = const()[name = string("op_1195_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1195_axis_0 = const()[name = string("op_1195_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1195_0, tensor<fp16, [1, 8, 128, 64]> var_1195_1 = split(axis = var_1195_axis_0, split_sizes = var_1195_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1195")];
            bool attn_weights_33_transpose_x_0 = const()[name = string("attn_weights_33_transpose_x_0"), val = bool(false)];
            bool attn_weights_33_transpose_y_0 = const()[name = string("attn_weights_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_33_cast_fp16 = matmul(transpose_x = attn_weights_33_transpose_x_0, transpose_y = attn_weights_33_transpose_y_0, x = var_1182_cast_fp16_0, y = var_1195_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1198_to_fp16 = const()[name = string("op_1198_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_35_cast_fp16 = mul(x = attn_weights_33_cast_fp16, y = var_1198_to_fp16)[name = string("attn_weights_35_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_37_cast_fp16 = add(x = attn_weights_35_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_37_cast_fp16")];
            int32 var_1202 = const()[name = string("op_1202"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_39_cast_fp16 = softmax(axis = var_1202, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1208_transpose_x_1 = const()[name = string("op_1208_transpose_x_1"), val = bool(true)];
            bool var_1208_transpose_y_1 = const()[name = string("op_1208_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1208_cast_fp16 = matmul(transpose_x = var_1208_transpose_x_1, transpose_y = var_1208_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1192_cast_fp16_0)[name = string("op_1208_cast_fp16")];
            bool attn_weights_41_transpose_x_0 = const()[name = string("attn_weights_41_transpose_x_0"), val = bool(false)];
            bool attn_weights_41_transpose_y_0 = const()[name = string("attn_weights_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_41_cast_fp16 = matmul(transpose_x = attn_weights_41_transpose_x_0, transpose_y = attn_weights_41_transpose_y_0, x = var_1182_cast_fp16_1, y = var_1195_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1210_to_fp16 = const()[name = string("op_1210_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_43_cast_fp16 = mul(x = attn_weights_41_cast_fp16, y = var_1210_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            int32 var_1214 = const()[name = string("op_1214"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_47_cast_fp16 = softmax(axis = var_1214, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")];
            bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(true)];
            bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_47_cast_fp16, y = var_1192_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1222 = const()[name = string("op_1222"), val = int32(1)];
            bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_19_cast_fp16 = concat(axis = var_1222, interleave = attn_output_19_interleave_0, values = (var_1208_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1226_perm_0 = const()[name = string("op_1226_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_1226_cast_fp16 = transpose(perm = var_1226_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1226_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1259_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            int32 var_1257 = const()[name = string("op_1257"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_21_cast_fp16 = concat(axis = var_1257, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1259_cast_fp16))[name = string("doubled_21_cast_fp16")];
            tensor<int32, [1]> out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_11_gamma_0_to_fp16 = const()[name = string("out_11_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368495360)))];
            fp16 var_1269_to_fp16 = const()[name = string("op_1269_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1269_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1280_split_sizes_0 = const()[name = string("op_1280_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1280_axis_0 = const()[name = string("op_1280_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1280_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1280_cast_fp16_1 = split(axis = var_1280_axis_0, split_sizes = var_1280_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1280_cast_fp16")];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_1297_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1297_cast_fp16")];
            tensor<int32, [2]> var_1303_strides_0 = const()[name = string("op_1303_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1303_pad_type_0 = const()[name = string("op_1303_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1303_pad_0 = const()[name = string("op_1303_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1303_dilations_0 = const()[name = string("op_1303_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1303_groups_0 = const()[name = string("op_1303_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_1303_cast_fp16 = conv(dilations = var_1303_dilations_0, groups = var_1303_groups_0, pad = var_1303_pad_0, pad_type = var_1303_pad_type_0, strides = var_1303_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("op_1303_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_29_cast_fp16 = mul(x = var_1297_cast_fp16, y = var_1303_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1321_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1321_cast_fp16")];
            int32 var_1319 = const()[name = string("op_1319"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_25_cast_fp16 = concat(axis = var_1319, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1321_cast_fp16))[name = string("doubled_25_cast_fp16")];
            tensor<int32, [1]> out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_13_gamma_0_to_fp16 = const()[name = string("out_13_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368503616)))];
            fp16 var_1331_to_fp16 = const()[name = string("op_1331_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1331_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1342_split_sizes_0 = const()[name = string("op_1342_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1342_axis_0 = const()[name = string("op_1342_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1342_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1342_cast_fp16_1 = split(axis = var_1342_axis_0, split_sizes = var_1342_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1342_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368511872)))];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = var_1342_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_1399_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_1406_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1406_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1410_cast_fp16 = mul(x = x_31_cast_fp16, y = var_308_cast_fp16)[name = string("op_1410_cast_fp16")];
            tensor<int32, [2]> var_1411_split_sizes_0 = const()[name = string("op_1411_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1411_axis_0 = const()[name = string("op_1411_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_1411_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1411_cast_fp16_1 = split(axis = var_1411_axis_0, split_sizes = var_1411_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1411_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_1413_cast_fp16 = mul(x = var_1411_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1413_cast_fp16")];
            int32 var_1415 = const()[name = string("op_1415"), val = int32(-2)];
            bool var_1416_interleave_0 = const()[name = string("op_1416_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1416_cast_fp16 = concat(axis = var_1415, interleave = var_1416_interleave_0, values = (var_1413_cast_fp16, var_1411_cast_fp16_0))[name = string("op_1416_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1417_cast_fp16 = mul(x = var_1416_cast_fp16, y = var_315_cast_fp16)[name = string("op_1417_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_21_cast_fp16 = add(x = var_1410_cast_fp16, y = var_1417_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1423_cast_fp16 = mul(x = var_1399_cast_fp16, y = var_308_cast_fp16)[name = string("op_1423_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_1424_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = var_1399_cast_fp16)[name = string("op_1424_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_1426_cast_fp16 = mul(x = var_1424_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1426_cast_fp16")];
            int32 var_1428 = const()[name = string("op_1428"), val = int32(-2)];
            bool var_1429_interleave_0 = const()[name = string("op_1429_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1429_cast_fp16 = concat(axis = var_1428, interleave = var_1429_interleave_0, values = (var_1426_cast_fp16, var_1424_cast_fp16_0))[name = string("op_1429_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1430_cast_fp16 = mul(x = var_1429_cast_fp16, y = var_315_cast_fp16)[name = string("op_1430_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_35_cast_fp16 = add(x = var_1423_cast_fp16, y = var_1430_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_167")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_94)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_96_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_96 = read_state(input = key_cache)[name = string("coreml_update_state_96")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1406_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_95)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_97_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_97 = read_state(input = value_cache)[name = string("coreml_update_state_97")];
            tensor<int32, [4]> var_1500_begin_0 = const()[name = string("op_1500_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1500_end_0 = const()[name = string("op_1500_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1500_end_mask_0 = const()[name = string("op_1500_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1500_cast_fp16 = slice_by_index(begin = var_1500_begin_0, end = var_1500_end_0, end_mask = var_1500_end_mask_0, x = coreml_update_state_96)[name = string("op_1500_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1503_axis_0 = const()[name = string("op_1503_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_1 = split(axis = var_1503_axis_0, split_sizes = tile_6, x = var_1500_cast_fp16)[name = string("op_1503_cast_fp16")];
            tensor<int32, [4]> var_1510_begin_0 = const()[name = string("op_1510_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1510_end_0 = const()[name = string("op_1510_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1510_end_mask_0 = const()[name = string("op_1510_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1510_cast_fp16 = slice_by_index(begin = var_1510_begin_0, end = var_1510_end_0, end_mask = var_1510_end_mask_0, x = coreml_update_state_97)[name = string("op_1510_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1513_axis_0 = const()[name = string("op_1513_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_1 = split(axis = var_1513_axis_0, split_sizes = tile_7, x = var_1510_cast_fp16)[name = string("op_1513_cast_fp16")];
            tensor<int32, [2]> var_1516_split_sizes_0 = const()[name = string("op_1516_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1516_axis_0 = const()[name = string("op_1516_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1516_0, tensor<fp16, [1, 8, 128, 64]> var_1516_1 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1516")];
            bool attn_weights_49_transpose_x_0 = const()[name = string("attn_weights_49_transpose_x_0"), val = bool(false)];
            bool attn_weights_49_transpose_y_0 = const()[name = string("attn_weights_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_0, transpose_y = attn_weights_49_transpose_y_0, x = var_1503_cast_fp16_0, y = var_1516_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_1519_to_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_53_cast_fp16 = add(x = attn_weights_51_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_53_cast_fp16")];
            int32 var_1523 = const()[name = string("op_1523"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_55_cast_fp16 = softmax(axis = var_1523, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1529_transpose_x_1 = const()[name = string("op_1529_transpose_x_1"), val = bool(true)];
            bool var_1529_transpose_y_1 = const()[name = string("op_1529_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1529_cast_fp16 = matmul(transpose_x = var_1529_transpose_x_1, transpose_y = var_1529_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1513_cast_fp16_0)[name = string("op_1529_cast_fp16")];
            bool attn_weights_57_transpose_x_0 = const()[name = string("attn_weights_57_transpose_x_0"), val = bool(false)];
            bool attn_weights_57_transpose_y_0 = const()[name = string("attn_weights_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_57_cast_fp16 = matmul(transpose_x = attn_weights_57_transpose_x_0, transpose_y = attn_weights_57_transpose_y_0, x = var_1503_cast_fp16_1, y = var_1516_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1531_to_fp16 = const()[name = string("op_1531_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_59_cast_fp16 = mul(x = attn_weights_57_cast_fp16, y = var_1531_to_fp16)[name = string("attn_weights_59_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_61_cast_fp16 = add(x = attn_weights_59_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_61_cast_fp16")];
            int32 var_1535 = const()[name = string("op_1535"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_63_cast_fp16 = softmax(axis = var_1535, x = attn_weights_61_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(true)];
            bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_63_cast_fp16, y = var_1513_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1543 = const()[name = string("op_1543"), val = int32(1)];
            bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_27_cast_fp16 = concat(axis = var_1543, interleave = attn_output_27_interleave_0, values = (var_1529_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1547_perm_0 = const()[name = string("op_1547_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_1547_cast_fp16 = transpose(perm = var_1547_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1547_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369560512)))];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1580_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1580_cast_fp16")];
            int32 var_1578 = const()[name = string("op_1578"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_29_cast_fp16 = concat(axis = var_1578, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1580_cast_fp16))[name = string("doubled_29_cast_fp16")];
            tensor<int32, [1]> out_15_axes_0 = const()[name = string("out_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_15_gamma_0_to_fp16 = const()[name = string("out_15_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377949184)))];
            fp16 var_1590_to_fp16 = const()[name = string("op_1590_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_1590_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1601_split_sizes_0 = const()[name = string("op_1601_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1601_axis_0 = const()[name = string("op_1601_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1601_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1601_cast_fp16_1 = split(axis = var_1601_axis_0, split_sizes = var_1601_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1601_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_1618_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1618_cast_fp16")];
            tensor<int32, [2]> var_1624_strides_0 = const()[name = string("op_1624_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1624_pad_type_0 = const()[name = string("op_1624_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1624_pad_0 = const()[name = string("op_1624_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1624_dilations_0 = const()[name = string("op_1624_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1624_groups_0 = const()[name = string("op_1624_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_1624_cast_fp16 = conv(dilations = var_1624_dilations_0, groups = var_1624_groups_0, pad = var_1624_pad_0, pad_type = var_1624_pad_type_0, strides = var_1624_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("op_1624_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_39_cast_fp16 = mul(x = var_1618_cast_fp16, y = var_1624_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1642_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1642_cast_fp16")];
            int32 var_1640 = const()[name = string("op_1640"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_33_cast_fp16 = concat(axis = var_1640, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1642_cast_fp16))[name = string("doubled_33_cast_fp16")];
            tensor<int32, [1]> out_17_axes_0 = const()[name = string("out_17_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_17_gamma_0_to_fp16 = const()[name = string("out_17_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377957440)))];
            fp16 var_1652_to_fp16 = const()[name = string("op_1652_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1652_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1663_split_sizes_0 = const()[name = string("op_1663_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1663_axis_0 = const()[name = string("op_1663_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1663_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1663_cast_fp16_1 = split(axis = var_1663_axis_0, split_sizes = var_1663_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1663_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377965696)))];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = var_1663_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_1720_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1720_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_1727_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1731_cast_fp16 = mul(x = x_41_cast_fp16, y = var_308_cast_fp16)[name = string("op_1731_cast_fp16")];
            tensor<int32, [2]> var_1732_split_sizes_0 = const()[name = string("op_1732_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1732_axis_0 = const()[name = string("op_1732_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_1732_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1732_cast_fp16_1 = split(axis = var_1732_axis_0, split_sizes = var_1732_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1732_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_1734_cast_fp16 = mul(x = var_1732_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1734_cast_fp16")];
            int32 var_1736 = const()[name = string("op_1736"), val = int32(-2)];
            bool var_1737_interleave_0 = const()[name = string("op_1737_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1737_cast_fp16 = concat(axis = var_1736, interleave = var_1737_interleave_0, values = (var_1734_cast_fp16, var_1732_cast_fp16_0))[name = string("op_1737_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1738_cast_fp16 = mul(x = var_1737_cast_fp16, y = var_315_cast_fp16)[name = string("op_1738_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_27_cast_fp16 = add(x = var_1731_cast_fp16, y = var_1738_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1744_cast_fp16 = mul(x = var_1720_cast_fp16, y = var_308_cast_fp16)[name = string("op_1744_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_1745_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = var_1720_cast_fp16)[name = string("op_1745_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_1747_cast_fp16 = mul(x = var_1745_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1747_cast_fp16")];
            int32 var_1749 = const()[name = string("op_1749"), val = int32(-2)];
            bool var_1750_interleave_0 = const()[name = string("op_1750_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1750_cast_fp16 = concat(axis = var_1749, interleave = var_1750_interleave_0, values = (var_1747_cast_fp16, var_1745_cast_fp16_0))[name = string("op_1750_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1751_cast_fp16 = mul(x = var_1750_cast_fp16, y = var_315_cast_fp16)[name = string("op_1751_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_45_cast_fp16 = add(x = var_1744_cast_fp16, y = var_1751_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_96)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_98_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_98 = read_state(input = key_cache)[name = string("coreml_update_state_98")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1727_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_97)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_99_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_99 = read_state(input = value_cache)[name = string("coreml_update_state_99")];
            tensor<int32, [4]> var_1821_begin_0 = const()[name = string("op_1821_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1821_end_0 = const()[name = string("op_1821_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1821_end_mask_0 = const()[name = string("op_1821_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1821_cast_fp16 = slice_by_index(begin = var_1821_begin_0, end = var_1821_end_0, end_mask = var_1821_end_mask_0, x = coreml_update_state_98)[name = string("op_1821_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1824_axis_0 = const()[name = string("op_1824_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_1 = split(axis = var_1824_axis_0, split_sizes = tile_8, x = var_1821_cast_fp16)[name = string("op_1824_cast_fp16")];
            tensor<int32, [4]> var_1831_begin_0 = const()[name = string("op_1831_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1831_end_0 = const()[name = string("op_1831_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1831_end_mask_0 = const()[name = string("op_1831_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1831_cast_fp16 = slice_by_index(begin = var_1831_begin_0, end = var_1831_end_0, end_mask = var_1831_end_mask_0, x = coreml_update_state_99)[name = string("op_1831_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1834_axis_0 = const()[name = string("op_1834_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_1 = split(axis = var_1834_axis_0, split_sizes = tile_9, x = var_1831_cast_fp16)[name = string("op_1834_cast_fp16")];
            tensor<int32, [2]> var_1837_split_sizes_0 = const()[name = string("op_1837_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1837_axis_0 = const()[name = string("op_1837_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1837_0, tensor<fp16, [1, 8, 128, 64]> var_1837_1 = split(axis = var_1837_axis_0, split_sizes = var_1837_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1837")];
            bool attn_weights_65_transpose_x_0 = const()[name = string("attn_weights_65_transpose_x_0"), val = bool(false)];
            bool attn_weights_65_transpose_y_0 = const()[name = string("attn_weights_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_65_cast_fp16 = matmul(transpose_x = attn_weights_65_transpose_x_0, transpose_y = attn_weights_65_transpose_y_0, x = var_1824_cast_fp16_0, y = var_1837_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1840_to_fp16 = const()[name = string("op_1840_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_67_cast_fp16 = mul(x = attn_weights_65_cast_fp16, y = var_1840_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            int32 var_1844 = const()[name = string("op_1844"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_71_cast_fp16 = softmax(axis = var_1844, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1850_transpose_x_1 = const()[name = string("op_1850_transpose_x_1"), val = bool(true)];
            bool var_1850_transpose_y_1 = const()[name = string("op_1850_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1850_cast_fp16 = matmul(transpose_x = var_1850_transpose_x_1, transpose_y = var_1850_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1834_cast_fp16_0)[name = string("op_1850_cast_fp16")];
            bool attn_weights_73_transpose_x_0 = const()[name = string("attn_weights_73_transpose_x_0"), val = bool(false)];
            bool attn_weights_73_transpose_y_0 = const()[name = string("attn_weights_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_0, transpose_y = attn_weights_73_transpose_y_0, x = var_1824_cast_fp16_1, y = var_1837_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1852_to_fp16 = const()[name = string("op_1852_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1852_to_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_77_cast_fp16 = add(x = attn_weights_75_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_77_cast_fp16")];
            int32 var_1856 = const()[name = string("op_1856"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_79_cast_fp16 = softmax(axis = var_1856, x = attn_weights_77_cast_fp16)[name = string("attn_weights_79_cast_fp16")];
            bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(true)];
            bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_79_cast_fp16, y = var_1834_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1864 = const()[name = string("op_1864"), val = int32(1)];
            bool attn_output_35_interleave_0 = const()[name = string("attn_output_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_35_cast_fp16 = concat(axis = var_1864, interleave = attn_output_35_interleave_0, values = (var_1850_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1868_perm_0 = const()[name = string("op_1868_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_1868_cast_fp16 = transpose(perm = var_1868_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1868_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1901_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1901_cast_fp16")];
            int32 var_1899 = const()[name = string("op_1899"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_37_cast_fp16 = concat(axis = var_1899, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1901_cast_fp16))[name = string("doubled_37_cast_fp16")];
            tensor<int32, [1]> out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_19_gamma_0_to_fp16 = const()[name = string("out_19_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379014336)))];
            fp16 var_1911_to_fp16 = const()[name = string("op_1911_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1911_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1922_split_sizes_0 = const()[name = string("op_1922_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1922_axis_0 = const()[name = string("op_1922_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1922_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1922_cast_fp16_1 = split(axis = var_1922_axis_0, split_sizes = var_1922_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1922_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_1939_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<int32, [2]> var_1945_strides_0 = const()[name = string("op_1945_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1945_pad_type_0 = const()[name = string("op_1945_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1945_pad_0 = const()[name = string("op_1945_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1945_dilations_0 = const()[name = string("op_1945_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1945_groups_0 = const()[name = string("op_1945_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_1945_cast_fp16 = conv(dilations = var_1945_dilations_0, groups = var_1945_groups_0, pad = var_1945_pad_0, pad_type = var_1945_pad_type_0, strides = var_1945_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("op_1945_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_49_cast_fp16 = mul(x = var_1939_cast_fp16, y = var_1945_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_1963_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1963_cast_fp16")];
            int32 var_1961 = const()[name = string("op_1961"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_41_cast_fp16 = concat(axis = var_1961, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1963_cast_fp16))[name = string("doubled_41_cast_fp16")];
            tensor<int32, [1]> out_21_axes_0 = const()[name = string("out_21_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_21_gamma_0_to_fp16 = const()[name = string("out_21_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379022592)))];
            fp16 var_1973_to_fp16 = const()[name = string("op_1973_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1973_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_1984_split_sizes_0 = const()[name = string("op_1984_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1984_axis_0 = const()[name = string("op_1984_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1984_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1984_cast_fp16_1 = split(axis = var_1984_axis_0, split_sizes = var_1984_split_sizes_0, x = out_21_cast_fp16)[name = string("op_1984_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379030848)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_1984_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_2041_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2041_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_2048_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2048_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2052_cast_fp16 = mul(x = x_51_cast_fp16, y = var_308_cast_fp16)[name = string("op_2052_cast_fp16")];
            tensor<int32, [2]> var_2053_split_sizes_0 = const()[name = string("op_2053_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2053_axis_0 = const()[name = string("op_2053_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_2053_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_2053_cast_fp16_1 = split(axis = var_2053_axis_0, split_sizes = var_2053_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2053_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_2055_cast_fp16 = mul(x = var_2053_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2055_cast_fp16")];
            int32 var_2057 = const()[name = string("op_2057"), val = int32(-2)];
            bool var_2058_interleave_0 = const()[name = string("op_2058_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2058_cast_fp16 = concat(axis = var_2057, interleave = var_2058_interleave_0, values = (var_2055_cast_fp16, var_2053_cast_fp16_0))[name = string("op_2058_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2059_cast_fp16 = mul(x = var_2058_cast_fp16, y = var_315_cast_fp16)[name = string("op_2059_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_33_cast_fp16 = add(x = var_2052_cast_fp16, y = var_2059_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2065_cast_fp16 = mul(x = var_2041_cast_fp16, y = var_308_cast_fp16)[name = string("op_2065_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_2066_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = var_2041_cast_fp16)[name = string("op_2066_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_2068_cast_fp16 = mul(x = var_2066_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2068_cast_fp16")];
            int32 var_2070 = const()[name = string("op_2070"), val = int32(-2)];
            bool var_2071_interleave_0 = const()[name = string("op_2071_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2071_cast_fp16 = concat(axis = var_2070, interleave = var_2071_interleave_0, values = (var_2068_cast_fp16, var_2066_cast_fp16_0))[name = string("op_2071_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2072_cast_fp16 = mul(x = var_2071_cast_fp16, y = var_315_cast_fp16)[name = string("op_2072_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_55_cast_fp16 = add(x = var_2065_cast_fp16, y = var_2072_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_98)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_100_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_100 = read_state(input = key_cache)[name = string("coreml_update_state_100")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2048_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_99)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_101_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_101 = read_state(input = value_cache)[name = string("coreml_update_state_101")];
            tensor<int32, [4]> var_2142_begin_0 = const()[name = string("op_2142_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2142_end_0 = const()[name = string("op_2142_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2142_end_mask_0 = const()[name = string("op_2142_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2142_cast_fp16 = slice_by_index(begin = var_2142_begin_0, end = var_2142_end_0, end_mask = var_2142_end_mask_0, x = coreml_update_state_100)[name = string("op_2142_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2145_axis_0 = const()[name = string("op_2145_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_1 = split(axis = var_2145_axis_0, split_sizes = tile_10, x = var_2142_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [4]> var_2152_begin_0 = const()[name = string("op_2152_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2152_end_0 = const()[name = string("op_2152_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2152_end_mask_0 = const()[name = string("op_2152_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2152_cast_fp16 = slice_by_index(begin = var_2152_begin_0, end = var_2152_end_0, end_mask = var_2152_end_mask_0, x = coreml_update_state_101)[name = string("op_2152_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2155_axis_0 = const()[name = string("op_2155_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_1 = split(axis = var_2155_axis_0, split_sizes = tile_11, x = var_2152_cast_fp16)[name = string("op_2155_cast_fp16")];
            tensor<int32, [2]> var_2158_split_sizes_0 = const()[name = string("op_2158_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2158_axis_0 = const()[name = string("op_2158_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_2158_0, tensor<fp16, [1, 8, 128, 64]> var_2158_1 = split(axis = var_2158_axis_0, split_sizes = var_2158_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2158")];
            bool attn_weights_81_transpose_x_0 = const()[name = string("attn_weights_81_transpose_x_0"), val = bool(false)];
            bool attn_weights_81_transpose_y_0 = const()[name = string("attn_weights_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_81_cast_fp16 = matmul(transpose_x = attn_weights_81_transpose_x_0, transpose_y = attn_weights_81_transpose_y_0, x = var_2145_cast_fp16_0, y = var_2158_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2161_to_fp16 = const()[name = string("op_2161_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_83_cast_fp16 = mul(x = attn_weights_81_cast_fp16, y = var_2161_to_fp16)[name = string("attn_weights_83_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_85_cast_fp16 = add(x = attn_weights_83_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_85_cast_fp16")];
            int32 var_2165 = const()[name = string("op_2165"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_87_cast_fp16 = softmax(axis = var_2165, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2171_transpose_x_1 = const()[name = string("op_2171_transpose_x_1"), val = bool(true)];
            bool var_2171_transpose_y_1 = const()[name = string("op_2171_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_2171_cast_fp16 = matmul(transpose_x = var_2171_transpose_x_1, transpose_y = var_2171_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2155_cast_fp16_0)[name = string("op_2171_cast_fp16")];
            bool attn_weights_89_transpose_x_0 = const()[name = string("attn_weights_89_transpose_x_0"), val = bool(false)];
            bool attn_weights_89_transpose_y_0 = const()[name = string("attn_weights_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_89_cast_fp16 = matmul(transpose_x = attn_weights_89_transpose_x_0, transpose_y = attn_weights_89_transpose_y_0, x = var_2145_cast_fp16_1, y = var_2158_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2173_to_fp16 = const()[name = string("op_2173_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_91_cast_fp16 = mul(x = attn_weights_89_cast_fp16, y = var_2173_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            int32 var_2177 = const()[name = string("op_2177"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_95_cast_fp16 = softmax(axis = var_2177, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")];
            bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(true)];
            bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_95_cast_fp16, y = var_2155_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2185 = const()[name = string("op_2185"), val = int32(1)];
            bool attn_output_43_interleave_0 = const()[name = string("attn_output_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_43_cast_fp16 = concat(axis = var_2185, interleave = attn_output_43_interleave_0, values = (var_2171_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2189_perm_0 = const()[name = string("op_2189_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_2189_cast_fp16 = transpose(perm = var_2189_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2189_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_2222_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2222_cast_fp16")];
            int32 var_2220 = const()[name = string("op_2220"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_45_cast_fp16 = concat(axis = var_2220, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2222_cast_fp16))[name = string("doubled_45_cast_fp16")];
            tensor<int32, [1]> out_23_axes_0 = const()[name = string("out_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_23_gamma_0_to_fp16 = const()[name = string("out_23_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380079488)))];
            fp16 var_2232_to_fp16 = const()[name = string("op_2232_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_2232_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2243_split_sizes_0 = const()[name = string("op_2243_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2243_axis_0 = const()[name = string("op_2243_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2243_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2243_cast_fp16_1 = split(axis = var_2243_axis_0, split_sizes = var_2243_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2243_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_2260_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<int32, [2]> var_2266_strides_0 = const()[name = string("op_2266_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2266_pad_type_0 = const()[name = string("op_2266_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2266_pad_0 = const()[name = string("op_2266_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2266_dilations_0 = const()[name = string("op_2266_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2266_groups_0 = const()[name = string("op_2266_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_2266_cast_fp16 = conv(dilations = var_2266_dilations_0, groups = var_2266_groups_0, pad = var_2266_pad_0, pad_type = var_2266_pad_type_0, strides = var_2266_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("op_2266_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_59_cast_fp16 = mul(x = var_2260_cast_fp16, y = var_2266_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_2284_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2284_cast_fp16")];
            int32 var_2282 = const()[name = string("op_2282"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_49_cast_fp16 = concat(axis = var_2282, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2284_cast_fp16))[name = string("doubled_49_cast_fp16")];
            tensor<int32, [1]> out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_25_gamma_0_to_fp16 = const()[name = string("out_25_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380087744)))];
            fp16 var_2294_to_fp16 = const()[name = string("op_2294_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_2294_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2305_split_sizes_0 = const()[name = string("op_2305_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2305_axis_0 = const()[name = string("op_2305_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2305_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2305_cast_fp16_1 = split(axis = var_2305_axis_0, split_sizes = var_2305_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380096000)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2305_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_2362_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2362_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_2369_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2369_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2373_cast_fp16 = mul(x = x_61_cast_fp16, y = var_308_cast_fp16)[name = string("op_2373_cast_fp16")];
            tensor<int32, [2]> var_2374_split_sizes_0 = const()[name = string("op_2374_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2374_axis_0 = const()[name = string("op_2374_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_2374_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_2374_cast_fp16_1 = split(axis = var_2374_axis_0, split_sizes = var_2374_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2376_cast_fp16")];
            int32 var_2378 = const()[name = string("op_2378"), val = int32(-2)];
            bool var_2379_interleave_0 = const()[name = string("op_2379_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2379_cast_fp16 = concat(axis = var_2378, interleave = var_2379_interleave_0, values = (var_2376_cast_fp16, var_2374_cast_fp16_0))[name = string("op_2379_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2380_cast_fp16 = mul(x = var_2379_cast_fp16, y = var_315_cast_fp16)[name = string("op_2380_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_39_cast_fp16 = add(x = var_2373_cast_fp16, y = var_2380_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2386_cast_fp16 = mul(x = var_2362_cast_fp16, y = var_308_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_2387_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = var_2362_cast_fp16)[name = string("op_2387_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_2389_cast_fp16 = mul(x = var_2387_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2389_cast_fp16")];
            int32 var_2391 = const()[name = string("op_2391"), val = int32(-2)];
            bool var_2392_interleave_0 = const()[name = string("op_2392_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2392_cast_fp16 = concat(axis = var_2391, interleave = var_2392_interleave_0, values = (var_2389_cast_fp16, var_2387_cast_fp16_0))[name = string("op_2392_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2393_cast_fp16 = mul(x = var_2392_cast_fp16, y = var_315_cast_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_65_cast_fp16 = add(x = var_2386_cast_fp16, y = var_2393_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_100)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_102_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_102 = read_state(input = key_cache)[name = string("coreml_update_state_102")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2369_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_101)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_103_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_103 = read_state(input = value_cache)[name = string("coreml_update_state_103")];
            tensor<int32, [4]> var_2463_begin_0 = const()[name = string("op_2463_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2463_end_0 = const()[name = string("op_2463_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2463_end_mask_0 = const()[name = string("op_2463_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2463_cast_fp16 = slice_by_index(begin = var_2463_begin_0, end = var_2463_end_0, end_mask = var_2463_end_mask_0, x = coreml_update_state_102)[name = string("op_2463_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2466_axis_0 = const()[name = string("op_2466_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_1 = split(axis = var_2466_axis_0, split_sizes = tile_12, x = var_2463_cast_fp16)[name = string("op_2466_cast_fp16")];
            tensor<int32, [4]> var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, x = coreml_update_state_103)[name = string("op_2473_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2476_axis_0 = const()[name = string("op_2476_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_1 = split(axis = var_2476_axis_0, split_sizes = tile_13, x = var_2473_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<int32, [2]> var_2479_split_sizes_0 = const()[name = string("op_2479_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2479_axis_0 = const()[name = string("op_2479_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_2479_0, tensor<fp16, [1, 8, 128, 64]> var_2479_1 = split(axis = var_2479_axis_0, split_sizes = var_2479_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2479")];
            bool attn_weights_97_transpose_x_0 = const()[name = string("attn_weights_97_transpose_x_0"), val = bool(false)];
            bool attn_weights_97_transpose_y_0 = const()[name = string("attn_weights_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_0, transpose_y = attn_weights_97_transpose_y_0, x = var_2466_cast_fp16_0, y = var_2479_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2482_to_fp16 = const()[name = string("op_2482_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_2482_to_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_101_cast_fp16 = add(x = attn_weights_99_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_101_cast_fp16")];
            int32 var_2486 = const()[name = string("op_2486"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_103_cast_fp16 = softmax(axis = var_2486, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2492_transpose_x_1 = const()[name = string("op_2492_transpose_x_1"), val = bool(true)];
            bool var_2492_transpose_y_1 = const()[name = string("op_2492_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_2492_cast_fp16 = matmul(transpose_x = var_2492_transpose_x_1, transpose_y = var_2492_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2476_cast_fp16_0)[name = string("op_2492_cast_fp16")];
            bool attn_weights_105_transpose_x_0 = const()[name = string("attn_weights_105_transpose_x_0"), val = bool(false)];
            bool attn_weights_105_transpose_y_0 = const()[name = string("attn_weights_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_105_cast_fp16 = matmul(transpose_x = attn_weights_105_transpose_x_0, transpose_y = attn_weights_105_transpose_y_0, x = var_2466_cast_fp16_1, y = var_2479_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2494_to_fp16 = const()[name = string("op_2494_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_107_cast_fp16 = mul(x = attn_weights_105_cast_fp16, y = var_2494_to_fp16)[name = string("attn_weights_107_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_109_cast_fp16 = add(x = attn_weights_107_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_109_cast_fp16")];
            int32 var_2498 = const()[name = string("op_2498"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_111_cast_fp16 = softmax(axis = var_2498, x = attn_weights_109_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(true)];
            bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_111_cast_fp16, y = var_2476_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2506 = const()[name = string("op_2506"), val = int32(1)];
            bool attn_output_51_interleave_0 = const()[name = string("attn_output_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_51_cast_fp16 = concat(axis = var_2506, interleave = attn_output_51_interleave_0, values = (var_2492_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2510_perm_0 = const()[name = string("op_2510_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_2510_cast_fp16 = transpose(perm = var_2510_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2510_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_2543_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2543_cast_fp16")];
            int32 var_2541 = const()[name = string("op_2541"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_53_cast_fp16 = concat(axis = var_2541, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2543_cast_fp16))[name = string("doubled_53_cast_fp16")];
            tensor<int32, [1]> out_27_axes_0 = const()[name = string("out_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_27_gamma_0_to_fp16 = const()[name = string("out_27_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381144640)))];
            fp16 var_2553_to_fp16 = const()[name = string("op_2553_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_2553_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2564_split_sizes_0 = const()[name = string("op_2564_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2564_axis_0 = const()[name = string("op_2564_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2564_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2564_cast_fp16_1 = split(axis = var_2564_axis_0, split_sizes = var_2564_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_2581_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [2]> var_2587_strides_0 = const()[name = string("op_2587_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2587_pad_type_0 = const()[name = string("op_2587_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2587_pad_0 = const()[name = string("op_2587_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2587_dilations_0 = const()[name = string("op_2587_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2587_groups_0 = const()[name = string("op_2587_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_2587_cast_fp16 = conv(dilations = var_2587_dilations_0, groups = var_2587_groups_0, pad = var_2587_pad_0, pad_type = var_2587_pad_type_0, strides = var_2587_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("op_2587_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_69_cast_fp16 = mul(x = var_2581_cast_fp16, y = var_2587_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381152896)))];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_to_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_2605_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2605_cast_fp16")];
            int32 var_2603 = const()[name = string("op_2603"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_57_cast_fp16 = concat(axis = var_2603, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2605_cast_fp16))[name = string("doubled_57_cast_fp16")];
            tensor<int32, [1]> out_29_axes_0 = const()[name = string("out_29_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_29_gamma_0_to_fp16 = const()[name = string("out_29_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406318784)))];
            fp16 var_2615_to_fp16 = const()[name = string("op_2615_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_2615_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2626_split_sizes_0 = const()[name = string("op_2626_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2626_axis_0 = const()[name = string("op_2626_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2626_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2626_cast_fp16_1 = split(axis = var_2626_axis_0, split_sizes = var_2626_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2626_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406327040)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2626_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_2683_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_2690_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2690_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2694_cast_fp16 = mul(x = x_71_cast_fp16, y = var_308_cast_fp16)[name = string("op_2694_cast_fp16")];
            tensor<int32, [2]> var_2695_split_sizes_0 = const()[name = string("op_2695_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2695_axis_0 = const()[name = string("op_2695_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_2695_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_2695_cast_fp16_1 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2695_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_2697_cast_fp16 = mul(x = var_2695_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2697_cast_fp16")];
            int32 var_2699 = const()[name = string("op_2699"), val = int32(-2)];
            bool var_2700_interleave_0 = const()[name = string("op_2700_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2700_cast_fp16 = concat(axis = var_2699, interleave = var_2700_interleave_0, values = (var_2697_cast_fp16, var_2695_cast_fp16_0))[name = string("op_2700_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2701_cast_fp16 = mul(x = var_2700_cast_fp16, y = var_315_cast_fp16)[name = string("op_2701_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_45_cast_fp16 = add(x = var_2694_cast_fp16, y = var_2701_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2707_cast_fp16 = mul(x = var_2683_cast_fp16, y = var_308_cast_fp16)[name = string("op_2707_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_2708_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = var_2683_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_2710_cast_fp16 = mul(x = var_2708_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2710_cast_fp16")];
            int32 var_2712 = const()[name = string("op_2712"), val = int32(-2)];
            bool var_2713_interleave_0 = const()[name = string("op_2713_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2713_cast_fp16 = concat(axis = var_2712, interleave = var_2713_interleave_0, values = (var_2710_cast_fp16, var_2708_cast_fp16_0))[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2714_cast_fp16 = mul(x = var_2713_cast_fp16, y = var_315_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_75_cast_fp16 = add(x = var_2707_cast_fp16, y = var_2714_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_102)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_104_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_104 = read_state(input = key_cache)[name = string("coreml_update_state_104")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2690_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_103)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_105_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_105 = read_state(input = value_cache)[name = string("coreml_update_state_105")];
            tensor<int32, [4]> var_2784_begin_0 = const()[name = string("op_2784_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2784_end_0 = const()[name = string("op_2784_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2784_end_mask_0 = const()[name = string("op_2784_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2784_cast_fp16 = slice_by_index(begin = var_2784_begin_0, end = var_2784_end_0, end_mask = var_2784_end_mask_0, x = coreml_update_state_104)[name = string("op_2784_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2787_axis_0 = const()[name = string("op_2787_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_1 = split(axis = var_2787_axis_0, split_sizes = tile_14, x = var_2784_cast_fp16)[name = string("op_2787_cast_fp16")];
            tensor<int32, [4]> var_2794_begin_0 = const()[name = string("op_2794_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2794_end_0 = const()[name = string("op_2794_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2794_end_mask_0 = const()[name = string("op_2794_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2794_cast_fp16 = slice_by_index(begin = var_2794_begin_0, end = var_2794_end_0, end_mask = var_2794_end_mask_0, x = coreml_update_state_105)[name = string("op_2794_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2797_axis_0 = const()[name = string("op_2797_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_1 = split(axis = var_2797_axis_0, split_sizes = tile_15, x = var_2794_cast_fp16)[name = string("op_2797_cast_fp16")];
            tensor<int32, [2]> var_2800_split_sizes_0 = const()[name = string("op_2800_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2800_axis_0 = const()[name = string("op_2800_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_2800_0, tensor<fp16, [1, 8, 128, 64]> var_2800_1 = split(axis = var_2800_axis_0, split_sizes = var_2800_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2800")];
            bool attn_weights_113_transpose_x_0 = const()[name = string("attn_weights_113_transpose_x_0"), val = bool(false)];
            bool attn_weights_113_transpose_y_0 = const()[name = string("attn_weights_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_113_cast_fp16 = matmul(transpose_x = attn_weights_113_transpose_x_0, transpose_y = attn_weights_113_transpose_y_0, x = var_2787_cast_fp16_0, y = var_2800_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2803_to_fp16 = const()[name = string("op_2803_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_115_cast_fp16 = mul(x = attn_weights_113_cast_fp16, y = var_2803_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            int32 var_2807 = const()[name = string("op_2807"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_119_cast_fp16 = softmax(axis = var_2807, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2813_transpose_x_1 = const()[name = string("op_2813_transpose_x_1"), val = bool(true)];
            bool var_2813_transpose_y_1 = const()[name = string("op_2813_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_2813_cast_fp16 = matmul(transpose_x = var_2813_transpose_x_1, transpose_y = var_2813_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2797_cast_fp16_0)[name = string("op_2813_cast_fp16")];
            bool attn_weights_121_transpose_x_0 = const()[name = string("attn_weights_121_transpose_x_0"), val = bool(false)];
            bool attn_weights_121_transpose_y_0 = const()[name = string("attn_weights_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_0, transpose_y = attn_weights_121_transpose_y_0, x = var_2787_cast_fp16_1, y = var_2800_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2815_to_fp16 = const()[name = string("op_2815_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_2815_to_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_125_cast_fp16 = add(x = attn_weights_123_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_125_cast_fp16")];
            int32 var_2819 = const()[name = string("op_2819"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_127_cast_fp16 = softmax(axis = var_2819, x = attn_weights_125_cast_fp16)[name = string("attn_weights_127_cast_fp16")];
            bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(true)];
            bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_127_cast_fp16, y = var_2797_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2827 = const()[name = string("op_2827"), val = int32(1)];
            bool attn_output_59_interleave_0 = const()[name = string("attn_output_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_59_cast_fp16 = concat(axis = var_2827, interleave = attn_output_59_interleave_0, values = (var_2813_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2831_perm_0 = const()[name = string("op_2831_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_2831_cast_fp16 = transpose(perm = var_2831_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2831_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_2864_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2864_cast_fp16")];
            int32 var_2862 = const()[name = string("op_2862"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_61_cast_fp16 = concat(axis = var_2862, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2864_cast_fp16))[name = string("doubled_61_cast_fp16")];
            tensor<int32, [1]> out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_31_gamma_0_to_fp16 = const()[name = string("out_31_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407375680)))];
            fp16 var_2874_to_fp16 = const()[name = string("op_2874_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_2874_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2885_split_sizes_0 = const()[name = string("op_2885_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2885_axis_0 = const()[name = string("op_2885_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2885_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2885_cast_fp16_1 = split(axis = var_2885_axis_0, split_sizes = var_2885_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2885_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2885_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_2902_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2902_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407383936)))];
            tensor<int32, [2]> var_2908_strides_0 = const()[name = string("op_2908_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2908_pad_type_0 = const()[name = string("op_2908_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2908_pad_0 = const()[name = string("op_2908_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2908_dilations_0 = const()[name = string("op_2908_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2908_groups_0 = const()[name = string("op_2908_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_2908_cast_fp16 = conv(dilations = var_2908_dilations_0, groups = var_2908_groups_0, pad = var_2908_pad_0, pad_type = var_2908_pad_type_0, strides = var_2908_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2885_cast_fp16_0)[name = string("op_2908_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_79_cast_fp16 = mul(x = var_2902_cast_fp16, y = var_2908_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432549824)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_2926_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2926_cast_fp16")];
            int32 var_2924 = const()[name = string("op_2924"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_65_cast_fp16 = concat(axis = var_2924, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2926_cast_fp16))[name = string("doubled_65_cast_fp16")];
            tensor<int32, [1]> out_33_axes_0 = const()[name = string("out_33_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_33_gamma_0_to_fp16 = const()[name = string("out_33_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457715712)))];
            fp16 var_2936_to_fp16 = const()[name = string("op_2936_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_2936_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2947_split_sizes_0 = const()[name = string("op_2947_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2947_axis_0 = const()[name = string("op_2947_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2947_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2947_cast_fp16_1 = split(axis = var_2947_axis_0, split_sizes = var_2947_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2947_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457723968)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466112640)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467161280)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_3004_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3004_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_3011_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3011_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3015_cast_fp16 = mul(x = x_81_cast_fp16, y = var_308_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<int32, [2]> var_3016_split_sizes_0 = const()[name = string("op_3016_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3016_axis_0 = const()[name = string("op_3016_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_3016_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_3016_cast_fp16_1 = split(axis = var_3016_axis_0, split_sizes = var_3016_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3016_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_3018_cast_fp16 = mul(x = var_3016_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3018_cast_fp16")];
            int32 var_3020 = const()[name = string("op_3020"), val = int32(-2)];
            bool var_3021_interleave_0 = const()[name = string("op_3021_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_3021_cast_fp16 = concat(axis = var_3020, interleave = var_3021_interleave_0, values = (var_3018_cast_fp16, var_3016_cast_fp16_0))[name = string("op_3021_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3022_cast_fp16 = mul(x = var_3021_cast_fp16, y = var_315_cast_fp16)[name = string("op_3022_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_51_cast_fp16 = add(x = var_3015_cast_fp16, y = var_3022_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3028_cast_fp16 = mul(x = var_3004_cast_fp16, y = var_308_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_3029_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = var_3004_cast_fp16)[name = string("op_3029_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_3031_cast_fp16 = mul(x = var_3029_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3031_cast_fp16")];
            int32 var_3033 = const()[name = string("op_3033"), val = int32(-2)];
            bool var_3034_interleave_0 = const()[name = string("op_3034_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3034_cast_fp16 = concat(axis = var_3033, interleave = var_3034_interleave_0, values = (var_3031_cast_fp16, var_3029_cast_fp16_0))[name = string("op_3034_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3035_cast_fp16 = mul(x = var_3034_cast_fp16, y = var_315_cast_fp16)[name = string("op_3035_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_85_cast_fp16 = add(x = var_3028_cast_fp16, y = var_3035_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_104)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_106_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_106 = read_state(input = key_cache)[name = string("coreml_update_state_106")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3011_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_105)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_107_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_107 = read_state(input = value_cache)[name = string("coreml_update_state_107")];
            tensor<int32, [4]> var_3105_begin_0 = const()[name = string("op_3105_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3105_end_0 = const()[name = string("op_3105_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3105_end_mask_0 = const()[name = string("op_3105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3105_cast_fp16 = slice_by_index(begin = var_3105_begin_0, end = var_3105_end_0, end_mask = var_3105_end_mask_0, x = coreml_update_state_106)[name = string("op_3105_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3108_axis_0 = const()[name = string("op_3108_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_1 = split(axis = var_3108_axis_0, split_sizes = tile_16, x = var_3105_cast_fp16)[name = string("op_3108_cast_fp16")];
            tensor<int32, [4]> var_3115_begin_0 = const()[name = string("op_3115_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3115_end_0 = const()[name = string("op_3115_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3115_end_mask_0 = const()[name = string("op_3115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3115_cast_fp16 = slice_by_index(begin = var_3115_begin_0, end = var_3115_end_0, end_mask = var_3115_end_mask_0, x = coreml_update_state_107)[name = string("op_3115_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3118_axis_0 = const()[name = string("op_3118_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_1 = split(axis = var_3118_axis_0, split_sizes = tile_17, x = var_3115_cast_fp16)[name = string("op_3118_cast_fp16")];
            tensor<int32, [2]> var_3121_split_sizes_0 = const()[name = string("op_3121_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3121_axis_0 = const()[name = string("op_3121_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_3121_0, tensor<fp16, [1, 8, 128, 64]> var_3121_1 = split(axis = var_3121_axis_0, split_sizes = var_3121_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3121")];
            bool attn_weights_129_transpose_x_0 = const()[name = string("attn_weights_129_transpose_x_0"), val = bool(false)];
            bool attn_weights_129_transpose_y_0 = const()[name = string("attn_weights_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_129_cast_fp16 = matmul(transpose_x = attn_weights_129_transpose_x_0, transpose_y = attn_weights_129_transpose_y_0, x = var_3108_cast_fp16_0, y = var_3121_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3124_to_fp16 = const()[name = string("op_3124_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_131_cast_fp16 = mul(x = attn_weights_129_cast_fp16, y = var_3124_to_fp16)[name = string("attn_weights_131_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_133_cast_fp16 = add(x = attn_weights_131_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_133_cast_fp16")];
            int32 var_3128 = const()[name = string("op_3128"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_135_cast_fp16 = softmax(axis = var_3128, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3134_transpose_x_1 = const()[name = string("op_3134_transpose_x_1"), val = bool(true)];
            bool var_3134_transpose_y_1 = const()[name = string("op_3134_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_3134_cast_fp16 = matmul(transpose_x = var_3134_transpose_x_1, transpose_y = var_3134_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3118_cast_fp16_0)[name = string("op_3134_cast_fp16")];
            bool attn_weights_137_transpose_x_0 = const()[name = string("attn_weights_137_transpose_x_0"), val = bool(false)];
            bool attn_weights_137_transpose_y_0 = const()[name = string("attn_weights_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_137_cast_fp16 = matmul(transpose_x = attn_weights_137_transpose_x_0, transpose_y = attn_weights_137_transpose_y_0, x = var_3108_cast_fp16_1, y = var_3121_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3136_to_fp16 = const()[name = string("op_3136_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_139_cast_fp16 = mul(x = attn_weights_137_cast_fp16, y = var_3136_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            int32 var_3140 = const()[name = string("op_3140"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_cast_fp16 = softmax(axis = var_3140, x = attn_weights_141_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(true)];
            bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_cast_fp16, y = var_3118_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3148 = const()[name = string("op_3148"), val = int32(1)];
            bool attn_output_67_interleave_0 = const()[name = string("attn_output_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_67_cast_fp16 = concat(axis = var_3148, interleave = attn_output_67_interleave_0, values = (var_3134_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3152_perm_0 = const()[name = string("op_3152_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_3152_cast_fp16 = transpose(perm = var_3152_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_cast_fp16 = reshape(shape = concat_107x, x = var_3152_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468209920)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_3185_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3185_cast_fp16")];
            int32 var_3183 = const()[name = string("op_3183"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_69_cast_fp16 = concat(axis = var_3183, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3185_cast_fp16))[name = string("doubled_69_cast_fp16")];
            tensor<int32, [1]> out_35_axes_0 = const()[name = string("out_35_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_35_gamma_0_to_fp16 = const()[name = string("out_35_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476598592)))];
            fp16 var_3195_to_fp16 = const()[name = string("op_3195_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_3195_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3206_split_sizes_0 = const()[name = string("op_3206_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3206_axis_0 = const()[name = string("op_3206_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3206_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3206_cast_fp16_1 = split(axis = var_3206_axis_0, split_sizes = var_3206_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3206_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476606848)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_8_mlp_gate_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_3223_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3223_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501772736)))];
            tensor<int32, [2]> var_3229_strides_0 = const()[name = string("op_3229_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3229_pad_type_0 = const()[name = string("op_3229_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3229_pad_0 = const()[name = string("op_3229_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3229_dilations_0 = const()[name = string("op_3229_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3229_groups_0 = const()[name = string("op_3229_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_3229_cast_fp16 = conv(dilations = var_3229_dilations_0, groups = var_3229_groups_0, pad = var_3229_pad_0, pad_type = var_3229_pad_type_0, strides = var_3229_strides_0, weight = layers_8_mlp_up_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("op_3229_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_cast_fp16 = mul(x = var_3223_cast_fp16, y = var_3229_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526938624)))];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_3247_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3247_cast_fp16")];
            int32 var_3245 = const()[name = string("op_3245"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_73_cast_fp16 = concat(axis = var_3245, interleave = doubled_73_interleave_0, values = (hidden_states_cast_fp16, var_3247_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_axes_0 = const()[name = string("out_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_gamma_0_to_fp16 = const()[name = string("out_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552104512)))];
            fp16 var_3257_to_fp16 = const()[name = string("op_3257_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3257_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3268_split_sizes_0 = const()[name = string("op_3268_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3268_axis_0 = const()[name = string("op_3268_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states, tensor<fp16, [1, 2048, 1, 64]> var_3268_cast_fp16_1 = split(axis = var_3268_axis_0, split_sizes = var_3268_split_sizes_0, x = out_cast_fp16)[name = string("op_3268_cast_fp16")];
        } -> (hidden_states);
    func length_8<ios18>(tensor<fp32, [1, 2048, 1, 8]> inputs_embeds, state<tensor<fp16, [9, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [8]> position_index_seed, state<tensor<fp16, [9, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725952))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17321280))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17308928))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17327488))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29922816))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29910464))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29929024))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42516160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42512000))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42518272))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243840))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243264))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47244160))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51442688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51438528))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444800))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64040128))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64027776))))[name = string("layers_1_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64046336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76633472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76629312))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76635584))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80834112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80829952))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80836224))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361152))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81360576))))[name = string("layers_2_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361472))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85560000))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85555840))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85562112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98145088))))[name = string("layers_2_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110758976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110746624))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110765184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123352320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123348160))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123354432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127552960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127548800))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127555072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128079424))))[name = string("layers_3_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080320))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140675648))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140663296))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140681856))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153277184))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153264832))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153283392))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165870528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165866368))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165872640))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170071168))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170067008))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170073280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170597632))))[name = string("layers_4_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174797056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174792896))))[name = string("layers_4_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174799168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187394496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187382144))))[name = string("layers_4_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187400704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199996032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199983680))))[name = string("layers_4_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_4_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200002240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212589376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212585216))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212591488))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216790016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216785856))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216792128))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317056))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217316480))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317376))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221515904))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221511744))))[name = string("layers_5_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221518016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234113344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234100992))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234119552))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246714880))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246702528))))[name = string("layers_5_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_5_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246721088))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259308224))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259304064))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259310336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263508864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263504704))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263510976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035328))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264036224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268234752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268230592))))[name = string("layers_6_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268236864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280832192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280819840))))[name = string("layers_6_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280838400))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293433728))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293421376))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293439936))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297638464))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297634304))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297640576))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165504))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298164928))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165824))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302364352))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302360192))))[name = string("layers_7_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302366464))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314961792))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314949440))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(8)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [8]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_280 = const()[name = string("op_280"), val = int32(0)];
            bool var_282_exclusive_0 = const()[name = string("op_282_exclusive_0"), val = bool(false)];
            bool var_282_reverse_0 = const()[name = string("op_282_reverse_0"), val = bool(false)];
            tensor<fp16, [8]> var_282_cast_fp16 = cumsum(axis = var_280, exclusive = var_282_exclusive_0, reverse = var_282_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_282_cast_fp16")];
            fp16 var_284_promoted_to_fp16 = const()[name = string("op_284_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [8]> position_offsets_cast_fp16 = sub(x = var_282_cast_fp16, y = var_284_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8]> var_287_cast_fp16 = expand_dims(axes = var_287_axes_0, x = position_offsets_cast_fp16)[name = string("op_287_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_11")];
            tensor<fp16, [1, 8]> position_ids_1_cast_fp16 = add(x = var_287_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 8]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_10")];
            tensor<bool, [1, 8]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 8]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 8]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314968000)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_4_y_0 = const()[name = string("greater_equal_4_y_0"), val = int32(0)];
            tensor<bool, [1, 8]> greater_equal_4 = greater_equal(x = select_0, y = greater_equal_4_y_0)[name = string("greater_equal_4")];
            int32 slice_by_index_4 = const()[name = string("slice_by_index_4"), val = int32(32768)];
            tensor<int32, [1, 8]> add_4 = add(x = select_0, y = slice_by_index_4)[name = string("add_4")];
            tensor<int32, [1, 8]> select_4 = select(a = select_0, b = add_4, cond = greater_equal_4)[name = string("select_4")];
            int32 cos_1_cast_fp16_axis_2 = const()[name = string("cos_1_cast_fp16_axis_2"), val = int32(0)];
            tensor<fp16, [1, 8, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_2, batch_dims = cos_1_batch_dims_0, indices = select_4, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323356672)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_2 = const()[name = string("sin_1_cast_fp16_axis_2"), val = int32(0)];
            tensor<fp16, [1, 8, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_2, batch_dims = sin_1_batch_dims_0, indices = select_4, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_306_perm_0 = const()[name = string("op_306_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_308_axes_0 = const()[name = string("op_308_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 8]> var_306_cast_fp16 = transpose(perm = var_306_perm_0, x = cos_1_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 1, 128, 8]> var_308_cast_fp16 = expand_dims(axes = var_308_axes_0, x = var_306_cast_fp16)[name = string("op_308_cast_fp16")];
            tensor<int32, [3]> var_313_perm_0 = const()[name = string("op_313_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 8]> var_313_cast_fp16 = transpose(perm = var_313_perm_0, x = sin_1_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 1, 128, 8]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = var_313_cast_fp16)[name = string("op_315_cast_fp16")];
            tensor<int32, [1]> var_334_axes_0 = const()[name = string("op_334_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 8, 1]> var_334 = expand_dims(axes = var_334_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_334")];
            tensor<int32, [1, 1, 2048]> var_327 = const()[name = string("op_327"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331745344)))];
            tensor<bool, [1, 8, 2048]> var_335 = greater(x = var_327, y = var_334)[name = string("op_335")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_342_axes_0 = const()[name = string("op_342_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 8, 2048]> var_335_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_335)[name = string("cast_9")];
            tensor<fp16, [1, 1, 8, 2048]> var_342_cast_fp16 = expand_dims(axes = var_342_axes_0, x = var_335_to_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_346_promoted_to_fp16 = const()[name = string("op_346_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 8]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_342_cast_fp16)[name = string("transpose_87")];
            tensor<bool, [1, 1, 2048, 8]> var_347_cast_fp16 = equal(x = mask_cast_fp16, y = var_346_promoted_to_fp16)[name = string("op_347_cast_fp16")];
            fp16 var_348_to_fp16 = const()[name = string("op_348_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 8]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_348_to_fp16, cond = var_347_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_8")];
            tensor<fp16, [1, 2048, 1, 8]> var_358_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_358_cast_fp16")];
            int32 var_356 = const()[name = string("op_356"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_1_cast_fp16 = concat(axis = var_356, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_358_cast_fp16))[name = string("doubled_1_cast_fp16")];
            tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_1_gamma_0_to_fp16 = const()[name = string("out_1_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331753600)))];
            fp16 var_368_to_fp16 = const()[name = string("op_368_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_368_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_379_split_sizes_0 = const()[name = string("op_379_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_379_axis_0 = const()[name = string("op_379_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_379_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_379_cast_fp16_1 = split(axis = var_379_axis_0, split_sizes = var_379_split_sizes_0, x = out_1_cast_fp16)[name = string("op_379_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331761856)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_379_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_436_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_436_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_443_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_443_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_447_cast_fp16 = mul(x = x_1_cast_fp16, y = var_308_cast_fp16)[name = string("op_447_cast_fp16")];
            tensor<int32, [2]> var_448_split_sizes_0 = const()[name = string("op_448_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_448_axis_0 = const()[name = string("op_448_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_448_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_448_cast_fp16_1 = split(axis = var_448_axis_0, split_sizes = var_448_split_sizes_0, x = x_1_cast_fp16)[name = string("op_448_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_450_cast_fp16 = mul(x = var_448_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_450_cast_fp16")];
            int32 var_452 = const()[name = string("op_452"), val = int32(-2)];
            bool var_453_interleave_0 = const()[name = string("op_453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_453_cast_fp16 = concat(axis = var_452, interleave = var_453_interleave_0, values = (var_450_cast_fp16, var_448_cast_fp16_0))[name = string("op_453_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_454_cast_fp16 = mul(x = var_453_cast_fp16, y = var_315_cast_fp16)[name = string("op_454_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_3_cast_fp16 = add(x = var_447_cast_fp16, y = var_454_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_460_cast_fp16 = mul(x = var_436_cast_fp16, y = var_308_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [2]> var_461_split_sizes_0 = const()[name = string("op_461_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_461_axis_0 = const()[name = string("op_461_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_461_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_461_cast_fp16_1 = split(axis = var_461_axis_0, split_sizes = var_461_split_sizes_0, x = var_436_cast_fp16)[name = string("op_461_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_463_cast_fp16 = mul(x = var_461_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_463_cast_fp16")];
            int32 var_465 = const()[name = string("op_465"), val = int32(-2)];
            bool var_466_interleave_0 = const()[name = string("op_466_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_466_cast_fp16 = concat(axis = var_465, interleave = var_466_interleave_0, values = (var_463_cast_fp16, var_461_cast_fp16_0))[name = string("op_466_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_467_cast_fp16 = mul(x = var_466_cast_fp16, y = var_315_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_5_cast_fp16 = add(x = var_460_cast_fp16, y = var_467_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_36_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_36 = read_state(input = key_cache)[name = string("coreml_update_state_36")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_443_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_37_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_37 = read_state(input = value_cache)[name = string("coreml_update_state_37")];
            tensor<int32, [4]> var_537_begin_0 = const()[name = string("op_537_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_537_end_0 = const()[name = string("op_537_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_537_end_mask_0 = const()[name = string("op_537_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_537_cast_fp16 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = coreml_update_state_36)[name = string("op_537_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_540_axis_0 = const()[name = string("op_540_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_1 = split(axis = var_540_axis_0, split_sizes = tile_0, x = var_537_cast_fp16)[name = string("op_540_cast_fp16")];
            tensor<int32, [4]> var_547_begin_0 = const()[name = string("op_547_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_547_end_0 = const()[name = string("op_547_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_547_end_mask_0 = const()[name = string("op_547_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_547_cast_fp16 = slice_by_index(begin = var_547_begin_0, end = var_547_end_0, end_mask = var_547_end_mask_0, x = coreml_update_state_37)[name = string("op_547_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_550_axis_0 = const()[name = string("op_550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_1 = split(axis = var_550_axis_0, split_sizes = tile_1, x = var_547_cast_fp16)[name = string("op_550_cast_fp16")];
            tensor<int32, [2]> var_553_split_sizes_0 = const()[name = string("op_553_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_553_axis_0 = const()[name = string("op_553_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_553_0, tensor<fp16, [1, 8, 128, 8]> var_553_1 = split(axis = var_553_axis_0, split_sizes = var_553_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_553")];
            bool attn_weights_1_transpose_x_0 = const()[name = string("attn_weights_1_transpose_x_0"), val = bool(false)];
            bool attn_weights_1_transpose_y_0 = const()[name = string("attn_weights_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_540_cast_fp16_0, y = var_553_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_556_to_fp16 = const()[name = string("op_556_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_556_to_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_5_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
            int32 var_560 = const()[name = string("op_560"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_7_cast_fp16 = softmax(axis = var_560, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_566_transpose_x_1 = const()[name = string("op_566_transpose_x_1"), val = bool(true)];
            bool var_566_transpose_y_1 = const()[name = string("op_566_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_566_cast_fp16 = matmul(transpose_x = var_566_transpose_x_1, transpose_y = var_566_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_550_cast_fp16_0)[name = string("op_566_cast_fp16")];
            bool attn_weights_9_transpose_x_0 = const()[name = string("attn_weights_9_transpose_x_0"), val = bool(false)];
            bool attn_weights_9_transpose_y_0 = const()[name = string("attn_weights_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_9_cast_fp16 = matmul(transpose_x = attn_weights_9_transpose_x_0, transpose_y = attn_weights_9_transpose_y_0, x = var_540_cast_fp16_1, y = var_553_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_568_to_fp16 = const()[name = string("op_568_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_11_cast_fp16 = mul(x = attn_weights_9_cast_fp16, y = var_568_to_fp16)[name = string("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_13_cast_fp16 = add(x = attn_weights_11_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_13_cast_fp16")];
            int32 var_572 = const()[name = string("op_572"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_15_cast_fp16 = softmax(axis = var_572, x = attn_weights_13_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(true)];
            bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_15_cast_fp16, y = var_550_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_580 = const()[name = string("op_580"), val = int32(1)];
            bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_3_cast_fp16 = concat(axis = var_580, interleave = attn_output_3_interleave_0, values = (var_566_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_584_perm_0 = const()[name = string("op_584_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_584_cast_fp16 = transpose(perm = var_584_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_584_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332810496)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_617_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_617_cast_fp16")];
            int32 var_615 = const()[name = string("op_615"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_5_cast_fp16 = concat(axis = var_615, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_617_cast_fp16))[name = string("doubled_5_cast_fp16")];
            tensor<int32, [1]> out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_3_gamma_0_to_fp16 = const()[name = string("out_3_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341199168)))];
            fp16 var_627_to_fp16 = const()[name = string("op_627_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_627_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_638_split_sizes_0 = const()[name = string("op_638_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_638_axis_0 = const()[name = string("op_638_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_638_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_638_cast_fp16_1 = split(axis = var_638_axis_0, split_sizes = var_638_split_sizes_0, x = out_3_cast_fp16)[name = string("op_638_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_655_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_655_cast_fp16")];
            tensor<int32, [2]> var_661_strides_0 = const()[name = string("op_661_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_661_pad_type_0 = const()[name = string("op_661_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_661_pad_0 = const()[name = string("op_661_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_661_dilations_0 = const()[name = string("op_661_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_661_groups_0 = const()[name = string("op_661_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_661_cast_fp16 = conv(dilations = var_661_dilations_0, groups = var_661_groups_0, pad = var_661_pad_0, pad_type = var_661_pad_type_0, strides = var_661_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("op_661_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_9_cast_fp16 = mul(x = var_655_cast_fp16, y = var_661_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_679_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_679_cast_fp16")];
            int32 var_677 = const()[name = string("op_677"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_9_cast_fp16 = concat(axis = var_677, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_679_cast_fp16))[name = string("doubled_9_cast_fp16")];
            tensor<int32, [1]> out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_5_gamma_0_to_fp16 = const()[name = string("out_5_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341207424)))];
            fp16 var_689_to_fp16 = const()[name = string("op_689_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_689_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_700_split_sizes_0 = const()[name = string("op_700_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_700_axis_0 = const()[name = string("op_700_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_700_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_700_cast_fp16_1 = split(axis = var_700_axis_0, split_sizes = var_700_split_sizes_0, x = out_5_cast_fp16)[name = string("op_700_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341215680)))];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = var_700_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_757_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_757_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_764_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_764_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_768_cast_fp16 = mul(x = x_11_cast_fp16, y = var_308_cast_fp16)[name = string("op_768_cast_fp16")];
            tensor<int32, [2]> var_769_split_sizes_0 = const()[name = string("op_769_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_769_axis_0 = const()[name = string("op_769_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_769_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_769_cast_fp16_1 = split(axis = var_769_axis_0, split_sizes = var_769_split_sizes_0, x = x_11_cast_fp16)[name = string("op_769_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_771_cast_fp16 = mul(x = var_769_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_771_cast_fp16")];
            int32 var_773 = const()[name = string("op_773"), val = int32(-2)];
            bool var_774_interleave_0 = const()[name = string("op_774_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_774_cast_fp16 = concat(axis = var_773, interleave = var_774_interleave_0, values = (var_771_cast_fp16, var_769_cast_fp16_0))[name = string("op_774_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_775_cast_fp16 = mul(x = var_774_cast_fp16, y = var_315_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_9_cast_fp16 = add(x = var_768_cast_fp16, y = var_775_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_781_cast_fp16 = mul(x = var_757_cast_fp16, y = var_308_cast_fp16)[name = string("op_781_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_782_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = var_757_cast_fp16)[name = string("op_782_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_784_cast_fp16 = mul(x = var_782_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_784_cast_fp16")];
            int32 var_786 = const()[name = string("op_786"), val = int32(-2)];
            bool var_787_interleave_0 = const()[name = string("op_787_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_787_cast_fp16 = concat(axis = var_786, interleave = var_787_interleave_0, values = (var_784_cast_fp16, var_782_cast_fp16_0))[name = string("op_787_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_788_cast_fp16 = mul(x = var_787_cast_fp16, y = var_315_cast_fp16)[name = string("op_788_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_15_cast_fp16 = add(x = var_781_cast_fp16, y = var_788_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_36)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_38_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_38 = read_state(input = key_cache)[name = string("coreml_update_state_38")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_764_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_37)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_39_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_39 = read_state(input = value_cache)[name = string("coreml_update_state_39")];
            tensor<int32, [4]> var_858_begin_0 = const()[name = string("op_858_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_858_end_0 = const()[name = string("op_858_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_858_end_mask_0 = const()[name = string("op_858_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_858_cast_fp16 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = coreml_update_state_38)[name = string("op_858_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_861_axis_0 = const()[name = string("op_861_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_1 = split(axis = var_861_axis_0, split_sizes = tile_2, x = var_858_cast_fp16)[name = string("op_861_cast_fp16")];
            tensor<int32, [4]> var_868_begin_0 = const()[name = string("op_868_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_868_end_0 = const()[name = string("op_868_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_868_end_mask_0 = const()[name = string("op_868_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_868_cast_fp16 = slice_by_index(begin = var_868_begin_0, end = var_868_end_0, end_mask = var_868_end_mask_0, x = coreml_update_state_39)[name = string("op_868_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_871_axis_0 = const()[name = string("op_871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_1 = split(axis = var_871_axis_0, split_sizes = tile_3, x = var_868_cast_fp16)[name = string("op_871_cast_fp16")];
            tensor<int32, [2]> var_874_split_sizes_0 = const()[name = string("op_874_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_874_axis_0 = const()[name = string("op_874_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_874_0, tensor<fp16, [1, 8, 128, 8]> var_874_1 = split(axis = var_874_axis_0, split_sizes = var_874_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_874")];
            bool attn_weights_17_transpose_x_0 = const()[name = string("attn_weights_17_transpose_x_0"), val = bool(false)];
            bool attn_weights_17_transpose_y_0 = const()[name = string("attn_weights_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_17_cast_fp16 = matmul(transpose_x = attn_weights_17_transpose_x_0, transpose_y = attn_weights_17_transpose_y_0, x = var_861_cast_fp16_0, y = var_874_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_877_to_fp16 = const()[name = string("op_877_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_19_cast_fp16 = mul(x = attn_weights_17_cast_fp16, y = var_877_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            int32 var_881 = const()[name = string("op_881"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_23_cast_fp16 = softmax(axis = var_881, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_887_transpose_x_1 = const()[name = string("op_887_transpose_x_1"), val = bool(true)];
            bool var_887_transpose_y_1 = const()[name = string("op_887_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_887_cast_fp16 = matmul(transpose_x = var_887_transpose_x_1, transpose_y = var_887_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_871_cast_fp16_0)[name = string("op_887_cast_fp16")];
            bool attn_weights_25_transpose_x_0 = const()[name = string("attn_weights_25_transpose_x_0"), val = bool(false)];
            bool attn_weights_25_transpose_y_0 = const()[name = string("attn_weights_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_0, transpose_y = attn_weights_25_transpose_y_0, x = var_861_cast_fp16_1, y = var_874_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_889_to_fp16 = const()[name = string("op_889_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_889_to_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_29_cast_fp16 = add(x = attn_weights_27_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
            int32 var_893 = const()[name = string("op_893"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_31_cast_fp16 = softmax(axis = var_893, x = attn_weights_29_cast_fp16)[name = string("attn_weights_31_cast_fp16")];
            bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(true)];
            bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_31_cast_fp16, y = var_871_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_901 = const()[name = string("op_901"), val = int32(1)];
            bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_11_cast_fp16 = concat(axis = var_901, interleave = attn_output_11_interleave_0, values = (var_887_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_905_perm_0 = const()[name = string("op_905_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_905_cast_fp16 = transpose(perm = var_905_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_905_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_938_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_938_cast_fp16")];
            int32 var_936 = const()[name = string("op_936"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_13_cast_fp16 = concat(axis = var_936, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_938_cast_fp16))[name = string("doubled_13_cast_fp16")];
            tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_7_gamma_0_to_fp16 = const()[name = string("out_7_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342264320)))];
            fp16 var_948_to_fp16 = const()[name = string("op_948_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_948_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_959_split_sizes_0 = const()[name = string("op_959_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_959_axis_0 = const()[name = string("op_959_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_959_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_959_cast_fp16_1 = split(axis = var_959_axis_0, split_sizes = var_959_split_sizes_0, x = out_7_cast_fp16)[name = string("op_959_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_959_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_976_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_976_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342272576)))];
            tensor<int32, [2]> var_982_strides_0 = const()[name = string("op_982_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_982_pad_type_0 = const()[name = string("op_982_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_982_pad_0 = const()[name = string("op_982_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_982_dilations_0 = const()[name = string("op_982_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_982_groups_0 = const()[name = string("op_982_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_982_cast_fp16 = conv(dilations = var_982_dilations_0, groups = var_982_groups_0, pad = var_982_pad_0, pad_type = var_982_pad_type_0, strides = var_982_strides_0, weight = layers_1_mlp_up_proj_weight_to_fp16, x = var_959_cast_fp16_0)[name = string("op_982_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_19_cast_fp16 = mul(x = var_976_cast_fp16, y = var_982_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1000_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1000_cast_fp16")];
            int32 var_998 = const()[name = string("op_998"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_17_cast_fp16 = concat(axis = var_998, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1000_cast_fp16))[name = string("doubled_17_cast_fp16")];
            tensor<int32, [1]> out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_9_gamma_0_to_fp16 = const()[name = string("out_9_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367438464)))];
            fp16 var_1010_to_fp16 = const()[name = string("op_1010_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1010_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1021_split_sizes_0 = const()[name = string("op_1021_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1021_axis_0 = const()[name = string("op_1021_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1021_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1021_cast_fp16_1 = split(axis = var_1021_axis_0, split_sizes = var_1021_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1021_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367446720)))];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = var_1021_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_1078_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1078_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_1085_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1085_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1089_cast_fp16 = mul(x = x_21_cast_fp16, y = var_308_cast_fp16)[name = string("op_1089_cast_fp16")];
            tensor<int32, [2]> var_1090_split_sizes_0 = const()[name = string("op_1090_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1090_axis_0 = const()[name = string("op_1090_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1090_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1090_cast_fp16_1 = split(axis = var_1090_axis_0, split_sizes = var_1090_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1090_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_1092_cast_fp16 = mul(x = var_1090_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            int32 var_1094 = const()[name = string("op_1094"), val = int32(-2)];
            bool var_1095_interleave_0 = const()[name = string("op_1095_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1095_cast_fp16 = concat(axis = var_1094, interleave = var_1095_interleave_0, values = (var_1092_cast_fp16, var_1090_cast_fp16_0))[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1096_cast_fp16 = mul(x = var_1095_cast_fp16, y = var_315_cast_fp16)[name = string("op_1096_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_15_cast_fp16 = add(x = var_1089_cast_fp16, y = var_1096_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1102_cast_fp16 = mul(x = var_1078_cast_fp16, y = var_308_cast_fp16)[name = string("op_1102_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1103_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = var_1078_cast_fp16)[name = string("op_1103_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_1105_cast_fp16 = mul(x = var_1103_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1105_cast_fp16")];
            int32 var_1107 = const()[name = string("op_1107"), val = int32(-2)];
            bool var_1108_interleave_0 = const()[name = string("op_1108_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1108_cast_fp16 = concat(axis = var_1107, interleave = var_1108_interleave_0, values = (var_1105_cast_fp16, var_1103_cast_fp16_0))[name = string("op_1108_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1109_cast_fp16 = mul(x = var_1108_cast_fp16, y = var_315_cast_fp16)[name = string("op_1109_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_25_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1109_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_38)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_40_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_40 = read_state(input = key_cache)[name = string("coreml_update_state_40")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1085_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_39)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_41_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_41 = read_state(input = value_cache)[name = string("coreml_update_state_41")];
            tensor<int32, [4]> var_1179_begin_0 = const()[name = string("op_1179_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1179_end_0 = const()[name = string("op_1179_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1179_end_mask_0 = const()[name = string("op_1179_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1179_cast_fp16 = slice_by_index(begin = var_1179_begin_0, end = var_1179_end_0, end_mask = var_1179_end_mask_0, x = coreml_update_state_40)[name = string("op_1179_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1182_axis_0 = const()[name = string("op_1182_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_1 = split(axis = var_1182_axis_0, split_sizes = tile_4, x = var_1179_cast_fp16)[name = string("op_1182_cast_fp16")];
            tensor<int32, [4]> var_1189_begin_0 = const()[name = string("op_1189_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1189_end_0 = const()[name = string("op_1189_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1189_end_mask_0 = const()[name = string("op_1189_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1189_cast_fp16 = slice_by_index(begin = var_1189_begin_0, end = var_1189_end_0, end_mask = var_1189_end_mask_0, x = coreml_update_state_41)[name = string("op_1189_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1192_axis_0 = const()[name = string("op_1192_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_1 = split(axis = var_1192_axis_0, split_sizes = tile_5, x = var_1189_cast_fp16)[name = string("op_1192_cast_fp16")];
            tensor<int32, [2]> var_1195_split_sizes_0 = const()[name = string("op_1195_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1195_axis_0 = const()[name = string("op_1195_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1195_0, tensor<fp16, [1, 8, 128, 8]> var_1195_1 = split(axis = var_1195_axis_0, split_sizes = var_1195_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1195")];
            bool attn_weights_33_transpose_x_0 = const()[name = string("attn_weights_33_transpose_x_0"), val = bool(false)];
            bool attn_weights_33_transpose_y_0 = const()[name = string("attn_weights_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_33_cast_fp16 = matmul(transpose_x = attn_weights_33_transpose_x_0, transpose_y = attn_weights_33_transpose_y_0, x = var_1182_cast_fp16_0, y = var_1195_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1198_to_fp16 = const()[name = string("op_1198_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_35_cast_fp16 = mul(x = attn_weights_33_cast_fp16, y = var_1198_to_fp16)[name = string("attn_weights_35_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_37_cast_fp16 = add(x = attn_weights_35_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_37_cast_fp16")];
            int32 var_1202 = const()[name = string("op_1202"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_39_cast_fp16 = softmax(axis = var_1202, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1208_transpose_x_1 = const()[name = string("op_1208_transpose_x_1"), val = bool(true)];
            bool var_1208_transpose_y_1 = const()[name = string("op_1208_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1208_cast_fp16 = matmul(transpose_x = var_1208_transpose_x_1, transpose_y = var_1208_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1192_cast_fp16_0)[name = string("op_1208_cast_fp16")];
            bool attn_weights_41_transpose_x_0 = const()[name = string("attn_weights_41_transpose_x_0"), val = bool(false)];
            bool attn_weights_41_transpose_y_0 = const()[name = string("attn_weights_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_41_cast_fp16 = matmul(transpose_x = attn_weights_41_transpose_x_0, transpose_y = attn_weights_41_transpose_y_0, x = var_1182_cast_fp16_1, y = var_1195_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1210_to_fp16 = const()[name = string("op_1210_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_43_cast_fp16 = mul(x = attn_weights_41_cast_fp16, y = var_1210_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            int32 var_1214 = const()[name = string("op_1214"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_47_cast_fp16 = softmax(axis = var_1214, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")];
            bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(true)];
            bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_47_cast_fp16, y = var_1192_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1222 = const()[name = string("op_1222"), val = int32(1)];
            bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_19_cast_fp16 = concat(axis = var_1222, interleave = attn_output_19_interleave_0, values = (var_1208_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1226_perm_0 = const()[name = string("op_1226_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_1226_cast_fp16 = transpose(perm = var_1226_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1226_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1259_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            int32 var_1257 = const()[name = string("op_1257"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_21_cast_fp16 = concat(axis = var_1257, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1259_cast_fp16))[name = string("doubled_21_cast_fp16")];
            tensor<int32, [1]> out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_11_gamma_0_to_fp16 = const()[name = string("out_11_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368495360)))];
            fp16 var_1269_to_fp16 = const()[name = string("op_1269_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1269_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1280_split_sizes_0 = const()[name = string("op_1280_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1280_axis_0 = const()[name = string("op_1280_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1280_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1280_cast_fp16_1 = split(axis = var_1280_axis_0, split_sizes = var_1280_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1280_cast_fp16")];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1297_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1297_cast_fp16")];
            tensor<int32, [2]> var_1303_strides_0 = const()[name = string("op_1303_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1303_pad_type_0 = const()[name = string("op_1303_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1303_pad_0 = const()[name = string("op_1303_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1303_dilations_0 = const()[name = string("op_1303_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1303_groups_0 = const()[name = string("op_1303_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1303_cast_fp16 = conv(dilations = var_1303_dilations_0, groups = var_1303_groups_0, pad = var_1303_pad_0, pad_type = var_1303_pad_type_0, strides = var_1303_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("op_1303_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_29_cast_fp16 = mul(x = var_1297_cast_fp16, y = var_1303_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1321_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1321_cast_fp16")];
            int32 var_1319 = const()[name = string("op_1319"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_25_cast_fp16 = concat(axis = var_1319, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1321_cast_fp16))[name = string("doubled_25_cast_fp16")];
            tensor<int32, [1]> out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_13_gamma_0_to_fp16 = const()[name = string("out_13_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368503616)))];
            fp16 var_1331_to_fp16 = const()[name = string("op_1331_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1331_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1342_split_sizes_0 = const()[name = string("op_1342_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1342_axis_0 = const()[name = string("op_1342_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1342_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1342_cast_fp16_1 = split(axis = var_1342_axis_0, split_sizes = var_1342_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1342_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368511872)))];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = var_1342_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_1399_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_1406_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1406_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1410_cast_fp16 = mul(x = x_31_cast_fp16, y = var_308_cast_fp16)[name = string("op_1410_cast_fp16")];
            tensor<int32, [2]> var_1411_split_sizes_0 = const()[name = string("op_1411_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1411_axis_0 = const()[name = string("op_1411_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1411_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1411_cast_fp16_1 = split(axis = var_1411_axis_0, split_sizes = var_1411_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1411_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_1413_cast_fp16 = mul(x = var_1411_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1413_cast_fp16")];
            int32 var_1415 = const()[name = string("op_1415"), val = int32(-2)];
            bool var_1416_interleave_0 = const()[name = string("op_1416_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1416_cast_fp16 = concat(axis = var_1415, interleave = var_1416_interleave_0, values = (var_1413_cast_fp16, var_1411_cast_fp16_0))[name = string("op_1416_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1417_cast_fp16 = mul(x = var_1416_cast_fp16, y = var_315_cast_fp16)[name = string("op_1417_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_21_cast_fp16 = add(x = var_1410_cast_fp16, y = var_1417_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1423_cast_fp16 = mul(x = var_1399_cast_fp16, y = var_308_cast_fp16)[name = string("op_1423_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1424_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = var_1399_cast_fp16)[name = string("op_1424_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_1426_cast_fp16 = mul(x = var_1424_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1426_cast_fp16")];
            int32 var_1428 = const()[name = string("op_1428"), val = int32(-2)];
            bool var_1429_interleave_0 = const()[name = string("op_1429_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1429_cast_fp16 = concat(axis = var_1428, interleave = var_1429_interleave_0, values = (var_1426_cast_fp16, var_1424_cast_fp16_0))[name = string("op_1429_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1430_cast_fp16 = mul(x = var_1429_cast_fp16, y = var_315_cast_fp16)[name = string("op_1430_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_35_cast_fp16 = add(x = var_1423_cast_fp16, y = var_1430_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_40)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_42_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_42 = read_state(input = key_cache)[name = string("coreml_update_state_42")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1406_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_41)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_43_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_43 = read_state(input = value_cache)[name = string("coreml_update_state_43")];
            tensor<int32, [4]> var_1500_begin_0 = const()[name = string("op_1500_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1500_end_0 = const()[name = string("op_1500_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1500_end_mask_0 = const()[name = string("op_1500_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1500_cast_fp16 = slice_by_index(begin = var_1500_begin_0, end = var_1500_end_0, end_mask = var_1500_end_mask_0, x = coreml_update_state_42)[name = string("op_1500_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1503_axis_0 = const()[name = string("op_1503_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_1 = split(axis = var_1503_axis_0, split_sizes = tile_6, x = var_1500_cast_fp16)[name = string("op_1503_cast_fp16")];
            tensor<int32, [4]> var_1510_begin_0 = const()[name = string("op_1510_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1510_end_0 = const()[name = string("op_1510_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1510_end_mask_0 = const()[name = string("op_1510_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1510_cast_fp16 = slice_by_index(begin = var_1510_begin_0, end = var_1510_end_0, end_mask = var_1510_end_mask_0, x = coreml_update_state_43)[name = string("op_1510_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1513_axis_0 = const()[name = string("op_1513_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_1 = split(axis = var_1513_axis_0, split_sizes = tile_7, x = var_1510_cast_fp16)[name = string("op_1513_cast_fp16")];
            tensor<int32, [2]> var_1516_split_sizes_0 = const()[name = string("op_1516_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1516_axis_0 = const()[name = string("op_1516_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1516_0, tensor<fp16, [1, 8, 128, 8]> var_1516_1 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1516")];
            bool attn_weights_49_transpose_x_0 = const()[name = string("attn_weights_49_transpose_x_0"), val = bool(false)];
            bool attn_weights_49_transpose_y_0 = const()[name = string("attn_weights_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_0, transpose_y = attn_weights_49_transpose_y_0, x = var_1503_cast_fp16_0, y = var_1516_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_1519_to_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_53_cast_fp16 = add(x = attn_weights_51_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_53_cast_fp16")];
            int32 var_1523 = const()[name = string("op_1523"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_55_cast_fp16 = softmax(axis = var_1523, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1529_transpose_x_1 = const()[name = string("op_1529_transpose_x_1"), val = bool(true)];
            bool var_1529_transpose_y_1 = const()[name = string("op_1529_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1529_cast_fp16 = matmul(transpose_x = var_1529_transpose_x_1, transpose_y = var_1529_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1513_cast_fp16_0)[name = string("op_1529_cast_fp16")];
            bool attn_weights_57_transpose_x_0 = const()[name = string("attn_weights_57_transpose_x_0"), val = bool(false)];
            bool attn_weights_57_transpose_y_0 = const()[name = string("attn_weights_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_57_cast_fp16 = matmul(transpose_x = attn_weights_57_transpose_x_0, transpose_y = attn_weights_57_transpose_y_0, x = var_1503_cast_fp16_1, y = var_1516_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1531_to_fp16 = const()[name = string("op_1531_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_59_cast_fp16 = mul(x = attn_weights_57_cast_fp16, y = var_1531_to_fp16)[name = string("attn_weights_59_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_61_cast_fp16 = add(x = attn_weights_59_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_61_cast_fp16")];
            int32 var_1535 = const()[name = string("op_1535"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_63_cast_fp16 = softmax(axis = var_1535, x = attn_weights_61_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(true)];
            bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_63_cast_fp16, y = var_1513_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1543 = const()[name = string("op_1543"), val = int32(1)];
            bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_27_cast_fp16 = concat(axis = var_1543, interleave = attn_output_27_interleave_0, values = (var_1529_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1547_perm_0 = const()[name = string("op_1547_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_1547_cast_fp16 = transpose(perm = var_1547_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1547_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369560512)))];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1580_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1580_cast_fp16")];
            int32 var_1578 = const()[name = string("op_1578"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_29_cast_fp16 = concat(axis = var_1578, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1580_cast_fp16))[name = string("doubled_29_cast_fp16")];
            tensor<int32, [1]> out_15_axes_0 = const()[name = string("out_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_15_gamma_0_to_fp16 = const()[name = string("out_15_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377949184)))];
            fp16 var_1590_to_fp16 = const()[name = string("op_1590_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_1590_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1601_split_sizes_0 = const()[name = string("op_1601_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1601_axis_0 = const()[name = string("op_1601_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1601_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1601_cast_fp16_1 = split(axis = var_1601_axis_0, split_sizes = var_1601_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1601_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1618_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1618_cast_fp16")];
            tensor<int32, [2]> var_1624_strides_0 = const()[name = string("op_1624_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1624_pad_type_0 = const()[name = string("op_1624_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1624_pad_0 = const()[name = string("op_1624_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1624_dilations_0 = const()[name = string("op_1624_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1624_groups_0 = const()[name = string("op_1624_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1624_cast_fp16 = conv(dilations = var_1624_dilations_0, groups = var_1624_groups_0, pad = var_1624_pad_0, pad_type = var_1624_pad_type_0, strides = var_1624_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("op_1624_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_39_cast_fp16 = mul(x = var_1618_cast_fp16, y = var_1624_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1642_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1642_cast_fp16")];
            int32 var_1640 = const()[name = string("op_1640"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_33_cast_fp16 = concat(axis = var_1640, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1642_cast_fp16))[name = string("doubled_33_cast_fp16")];
            tensor<int32, [1]> out_17_axes_0 = const()[name = string("out_17_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_17_gamma_0_to_fp16 = const()[name = string("out_17_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377957440)))];
            fp16 var_1652_to_fp16 = const()[name = string("op_1652_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1652_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1663_split_sizes_0 = const()[name = string("op_1663_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1663_axis_0 = const()[name = string("op_1663_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1663_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1663_cast_fp16_1 = split(axis = var_1663_axis_0, split_sizes = var_1663_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1663_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377965696)))];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = var_1663_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_1720_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1720_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_1727_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1731_cast_fp16 = mul(x = x_41_cast_fp16, y = var_308_cast_fp16)[name = string("op_1731_cast_fp16")];
            tensor<int32, [2]> var_1732_split_sizes_0 = const()[name = string("op_1732_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1732_axis_0 = const()[name = string("op_1732_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1732_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1732_cast_fp16_1 = split(axis = var_1732_axis_0, split_sizes = var_1732_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1732_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_1734_cast_fp16 = mul(x = var_1732_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1734_cast_fp16")];
            int32 var_1736 = const()[name = string("op_1736"), val = int32(-2)];
            bool var_1737_interleave_0 = const()[name = string("op_1737_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1737_cast_fp16 = concat(axis = var_1736, interleave = var_1737_interleave_0, values = (var_1734_cast_fp16, var_1732_cast_fp16_0))[name = string("op_1737_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1738_cast_fp16 = mul(x = var_1737_cast_fp16, y = var_315_cast_fp16)[name = string("op_1738_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_27_cast_fp16 = add(x = var_1731_cast_fp16, y = var_1738_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1744_cast_fp16 = mul(x = var_1720_cast_fp16, y = var_308_cast_fp16)[name = string("op_1744_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1745_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = var_1720_cast_fp16)[name = string("op_1745_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_1747_cast_fp16 = mul(x = var_1745_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1747_cast_fp16")];
            int32 var_1749 = const()[name = string("op_1749"), val = int32(-2)];
            bool var_1750_interleave_0 = const()[name = string("op_1750_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1750_cast_fp16 = concat(axis = var_1749, interleave = var_1750_interleave_0, values = (var_1747_cast_fp16, var_1745_cast_fp16_0))[name = string("op_1750_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1751_cast_fp16 = mul(x = var_1750_cast_fp16, y = var_315_cast_fp16)[name = string("op_1751_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_45_cast_fp16 = add(x = var_1744_cast_fp16, y = var_1751_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_42)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_44_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_44 = read_state(input = key_cache)[name = string("coreml_update_state_44")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1727_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_43)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_45_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_45 = read_state(input = value_cache)[name = string("coreml_update_state_45")];
            tensor<int32, [4]> var_1821_begin_0 = const()[name = string("op_1821_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1821_end_0 = const()[name = string("op_1821_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1821_end_mask_0 = const()[name = string("op_1821_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1821_cast_fp16 = slice_by_index(begin = var_1821_begin_0, end = var_1821_end_0, end_mask = var_1821_end_mask_0, x = coreml_update_state_44)[name = string("op_1821_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1824_axis_0 = const()[name = string("op_1824_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_1 = split(axis = var_1824_axis_0, split_sizes = tile_8, x = var_1821_cast_fp16)[name = string("op_1824_cast_fp16")];
            tensor<int32, [4]> var_1831_begin_0 = const()[name = string("op_1831_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1831_end_0 = const()[name = string("op_1831_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1831_end_mask_0 = const()[name = string("op_1831_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1831_cast_fp16 = slice_by_index(begin = var_1831_begin_0, end = var_1831_end_0, end_mask = var_1831_end_mask_0, x = coreml_update_state_45)[name = string("op_1831_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1834_axis_0 = const()[name = string("op_1834_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_1 = split(axis = var_1834_axis_0, split_sizes = tile_9, x = var_1831_cast_fp16)[name = string("op_1834_cast_fp16")];
            tensor<int32, [2]> var_1837_split_sizes_0 = const()[name = string("op_1837_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1837_axis_0 = const()[name = string("op_1837_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1837_0, tensor<fp16, [1, 8, 128, 8]> var_1837_1 = split(axis = var_1837_axis_0, split_sizes = var_1837_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1837")];
            bool attn_weights_65_transpose_x_0 = const()[name = string("attn_weights_65_transpose_x_0"), val = bool(false)];
            bool attn_weights_65_transpose_y_0 = const()[name = string("attn_weights_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_65_cast_fp16 = matmul(transpose_x = attn_weights_65_transpose_x_0, transpose_y = attn_weights_65_transpose_y_0, x = var_1824_cast_fp16_0, y = var_1837_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1840_to_fp16 = const()[name = string("op_1840_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_67_cast_fp16 = mul(x = attn_weights_65_cast_fp16, y = var_1840_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            int32 var_1844 = const()[name = string("op_1844"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_71_cast_fp16 = softmax(axis = var_1844, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1850_transpose_x_1 = const()[name = string("op_1850_transpose_x_1"), val = bool(true)];
            bool var_1850_transpose_y_1 = const()[name = string("op_1850_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1850_cast_fp16 = matmul(transpose_x = var_1850_transpose_x_1, transpose_y = var_1850_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1834_cast_fp16_0)[name = string("op_1850_cast_fp16")];
            bool attn_weights_73_transpose_x_0 = const()[name = string("attn_weights_73_transpose_x_0"), val = bool(false)];
            bool attn_weights_73_transpose_y_0 = const()[name = string("attn_weights_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_0, transpose_y = attn_weights_73_transpose_y_0, x = var_1824_cast_fp16_1, y = var_1837_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1852_to_fp16 = const()[name = string("op_1852_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1852_to_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_77_cast_fp16 = add(x = attn_weights_75_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_77_cast_fp16")];
            int32 var_1856 = const()[name = string("op_1856"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_79_cast_fp16 = softmax(axis = var_1856, x = attn_weights_77_cast_fp16)[name = string("attn_weights_79_cast_fp16")];
            bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(true)];
            bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_79_cast_fp16, y = var_1834_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1864 = const()[name = string("op_1864"), val = int32(1)];
            bool attn_output_35_interleave_0 = const()[name = string("attn_output_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_35_cast_fp16 = concat(axis = var_1864, interleave = attn_output_35_interleave_0, values = (var_1850_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1868_perm_0 = const()[name = string("op_1868_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_1868_cast_fp16 = transpose(perm = var_1868_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1868_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1901_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1901_cast_fp16")];
            int32 var_1899 = const()[name = string("op_1899"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_37_cast_fp16 = concat(axis = var_1899, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1901_cast_fp16))[name = string("doubled_37_cast_fp16")];
            tensor<int32, [1]> out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_19_gamma_0_to_fp16 = const()[name = string("out_19_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379014336)))];
            fp16 var_1911_to_fp16 = const()[name = string("op_1911_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1911_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1922_split_sizes_0 = const()[name = string("op_1922_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1922_axis_0 = const()[name = string("op_1922_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1922_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1922_cast_fp16_1 = split(axis = var_1922_axis_0, split_sizes = var_1922_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1922_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1939_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<int32, [2]> var_1945_strides_0 = const()[name = string("op_1945_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1945_pad_type_0 = const()[name = string("op_1945_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1945_pad_0 = const()[name = string("op_1945_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1945_dilations_0 = const()[name = string("op_1945_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1945_groups_0 = const()[name = string("op_1945_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1945_cast_fp16 = conv(dilations = var_1945_dilations_0, groups = var_1945_groups_0, pad = var_1945_pad_0, pad_type = var_1945_pad_type_0, strides = var_1945_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("op_1945_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_49_cast_fp16 = mul(x = var_1939_cast_fp16, y = var_1945_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1963_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1963_cast_fp16")];
            int32 var_1961 = const()[name = string("op_1961"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_41_cast_fp16 = concat(axis = var_1961, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1963_cast_fp16))[name = string("doubled_41_cast_fp16")];
            tensor<int32, [1]> out_21_axes_0 = const()[name = string("out_21_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_21_gamma_0_to_fp16 = const()[name = string("out_21_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379022592)))];
            fp16 var_1973_to_fp16 = const()[name = string("op_1973_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1973_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_1984_split_sizes_0 = const()[name = string("op_1984_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1984_axis_0 = const()[name = string("op_1984_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1984_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1984_cast_fp16_1 = split(axis = var_1984_axis_0, split_sizes = var_1984_split_sizes_0, x = out_21_cast_fp16)[name = string("op_1984_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379030848)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_1984_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_2041_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2041_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_2048_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2048_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2052_cast_fp16 = mul(x = x_51_cast_fp16, y = var_308_cast_fp16)[name = string("op_2052_cast_fp16")];
            tensor<int32, [2]> var_2053_split_sizes_0 = const()[name = string("op_2053_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2053_axis_0 = const()[name = string("op_2053_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2053_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2053_cast_fp16_1 = split(axis = var_2053_axis_0, split_sizes = var_2053_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2053_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_2055_cast_fp16 = mul(x = var_2053_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2055_cast_fp16")];
            int32 var_2057 = const()[name = string("op_2057"), val = int32(-2)];
            bool var_2058_interleave_0 = const()[name = string("op_2058_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2058_cast_fp16 = concat(axis = var_2057, interleave = var_2058_interleave_0, values = (var_2055_cast_fp16, var_2053_cast_fp16_0))[name = string("op_2058_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2059_cast_fp16 = mul(x = var_2058_cast_fp16, y = var_315_cast_fp16)[name = string("op_2059_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_33_cast_fp16 = add(x = var_2052_cast_fp16, y = var_2059_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2065_cast_fp16 = mul(x = var_2041_cast_fp16, y = var_308_cast_fp16)[name = string("op_2065_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2066_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = var_2041_cast_fp16)[name = string("op_2066_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_2068_cast_fp16 = mul(x = var_2066_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2068_cast_fp16")];
            int32 var_2070 = const()[name = string("op_2070"), val = int32(-2)];
            bool var_2071_interleave_0 = const()[name = string("op_2071_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2071_cast_fp16 = concat(axis = var_2070, interleave = var_2071_interleave_0, values = (var_2068_cast_fp16, var_2066_cast_fp16_0))[name = string("op_2071_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2072_cast_fp16 = mul(x = var_2071_cast_fp16, y = var_315_cast_fp16)[name = string("op_2072_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_55_cast_fp16 = add(x = var_2065_cast_fp16, y = var_2072_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_44)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_46_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_46 = read_state(input = key_cache)[name = string("coreml_update_state_46")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2048_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_45)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_47_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_47 = read_state(input = value_cache)[name = string("coreml_update_state_47")];
            tensor<int32, [4]> var_2142_begin_0 = const()[name = string("op_2142_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2142_end_0 = const()[name = string("op_2142_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2142_end_mask_0 = const()[name = string("op_2142_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2142_cast_fp16 = slice_by_index(begin = var_2142_begin_0, end = var_2142_end_0, end_mask = var_2142_end_mask_0, x = coreml_update_state_46)[name = string("op_2142_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2145_axis_0 = const()[name = string("op_2145_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_1 = split(axis = var_2145_axis_0, split_sizes = tile_10, x = var_2142_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [4]> var_2152_begin_0 = const()[name = string("op_2152_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2152_end_0 = const()[name = string("op_2152_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2152_end_mask_0 = const()[name = string("op_2152_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2152_cast_fp16 = slice_by_index(begin = var_2152_begin_0, end = var_2152_end_0, end_mask = var_2152_end_mask_0, x = coreml_update_state_47)[name = string("op_2152_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2155_axis_0 = const()[name = string("op_2155_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_1 = split(axis = var_2155_axis_0, split_sizes = tile_11, x = var_2152_cast_fp16)[name = string("op_2155_cast_fp16")];
            tensor<int32, [2]> var_2158_split_sizes_0 = const()[name = string("op_2158_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2158_axis_0 = const()[name = string("op_2158_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2158_0, tensor<fp16, [1, 8, 128, 8]> var_2158_1 = split(axis = var_2158_axis_0, split_sizes = var_2158_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2158")];
            bool attn_weights_81_transpose_x_0 = const()[name = string("attn_weights_81_transpose_x_0"), val = bool(false)];
            bool attn_weights_81_transpose_y_0 = const()[name = string("attn_weights_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_81_cast_fp16 = matmul(transpose_x = attn_weights_81_transpose_x_0, transpose_y = attn_weights_81_transpose_y_0, x = var_2145_cast_fp16_0, y = var_2158_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2161_to_fp16 = const()[name = string("op_2161_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_83_cast_fp16 = mul(x = attn_weights_81_cast_fp16, y = var_2161_to_fp16)[name = string("attn_weights_83_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_85_cast_fp16 = add(x = attn_weights_83_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_85_cast_fp16")];
            int32 var_2165 = const()[name = string("op_2165"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_87_cast_fp16 = softmax(axis = var_2165, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2171_transpose_x_1 = const()[name = string("op_2171_transpose_x_1"), val = bool(true)];
            bool var_2171_transpose_y_1 = const()[name = string("op_2171_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2171_cast_fp16 = matmul(transpose_x = var_2171_transpose_x_1, transpose_y = var_2171_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2155_cast_fp16_0)[name = string("op_2171_cast_fp16")];
            bool attn_weights_89_transpose_x_0 = const()[name = string("attn_weights_89_transpose_x_0"), val = bool(false)];
            bool attn_weights_89_transpose_y_0 = const()[name = string("attn_weights_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_89_cast_fp16 = matmul(transpose_x = attn_weights_89_transpose_x_0, transpose_y = attn_weights_89_transpose_y_0, x = var_2145_cast_fp16_1, y = var_2158_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2173_to_fp16 = const()[name = string("op_2173_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_91_cast_fp16 = mul(x = attn_weights_89_cast_fp16, y = var_2173_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            int32 var_2177 = const()[name = string("op_2177"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_95_cast_fp16 = softmax(axis = var_2177, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")];
            bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(true)];
            bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_95_cast_fp16, y = var_2155_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2185 = const()[name = string("op_2185"), val = int32(1)];
            bool attn_output_43_interleave_0 = const()[name = string("attn_output_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_43_cast_fp16 = concat(axis = var_2185, interleave = attn_output_43_interleave_0, values = (var_2171_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2189_perm_0 = const()[name = string("op_2189_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_2189_cast_fp16 = transpose(perm = var_2189_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2189_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2222_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2222_cast_fp16")];
            int32 var_2220 = const()[name = string("op_2220"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_45_cast_fp16 = concat(axis = var_2220, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2222_cast_fp16))[name = string("doubled_45_cast_fp16")];
            tensor<int32, [1]> out_23_axes_0 = const()[name = string("out_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_23_gamma_0_to_fp16 = const()[name = string("out_23_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380079488)))];
            fp16 var_2232_to_fp16 = const()[name = string("op_2232_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_2232_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2243_split_sizes_0 = const()[name = string("op_2243_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2243_axis_0 = const()[name = string("op_2243_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2243_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2243_cast_fp16_1 = split(axis = var_2243_axis_0, split_sizes = var_2243_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2243_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2260_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<int32, [2]> var_2266_strides_0 = const()[name = string("op_2266_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2266_pad_type_0 = const()[name = string("op_2266_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2266_pad_0 = const()[name = string("op_2266_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2266_dilations_0 = const()[name = string("op_2266_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2266_groups_0 = const()[name = string("op_2266_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2266_cast_fp16 = conv(dilations = var_2266_dilations_0, groups = var_2266_groups_0, pad = var_2266_pad_0, pad_type = var_2266_pad_type_0, strides = var_2266_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("op_2266_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_59_cast_fp16 = mul(x = var_2260_cast_fp16, y = var_2266_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2284_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2284_cast_fp16")];
            int32 var_2282 = const()[name = string("op_2282"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_49_cast_fp16 = concat(axis = var_2282, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2284_cast_fp16))[name = string("doubled_49_cast_fp16")];
            tensor<int32, [1]> out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_25_gamma_0_to_fp16 = const()[name = string("out_25_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380087744)))];
            fp16 var_2294_to_fp16 = const()[name = string("op_2294_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_2294_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2305_split_sizes_0 = const()[name = string("op_2305_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2305_axis_0 = const()[name = string("op_2305_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2305_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2305_cast_fp16_1 = split(axis = var_2305_axis_0, split_sizes = var_2305_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380096000)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2305_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_2362_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2362_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_2369_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2369_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2373_cast_fp16 = mul(x = x_61_cast_fp16, y = var_308_cast_fp16)[name = string("op_2373_cast_fp16")];
            tensor<int32, [2]> var_2374_split_sizes_0 = const()[name = string("op_2374_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2374_axis_0 = const()[name = string("op_2374_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2374_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2374_cast_fp16_1 = split(axis = var_2374_axis_0, split_sizes = var_2374_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2376_cast_fp16")];
            int32 var_2378 = const()[name = string("op_2378"), val = int32(-2)];
            bool var_2379_interleave_0 = const()[name = string("op_2379_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2379_cast_fp16 = concat(axis = var_2378, interleave = var_2379_interleave_0, values = (var_2376_cast_fp16, var_2374_cast_fp16_0))[name = string("op_2379_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2380_cast_fp16 = mul(x = var_2379_cast_fp16, y = var_315_cast_fp16)[name = string("op_2380_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_39_cast_fp16 = add(x = var_2373_cast_fp16, y = var_2380_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2386_cast_fp16 = mul(x = var_2362_cast_fp16, y = var_308_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2387_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = var_2362_cast_fp16)[name = string("op_2387_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_2389_cast_fp16 = mul(x = var_2387_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2389_cast_fp16")];
            int32 var_2391 = const()[name = string("op_2391"), val = int32(-2)];
            bool var_2392_interleave_0 = const()[name = string("op_2392_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2392_cast_fp16 = concat(axis = var_2391, interleave = var_2392_interleave_0, values = (var_2389_cast_fp16, var_2387_cast_fp16_0))[name = string("op_2392_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2393_cast_fp16 = mul(x = var_2392_cast_fp16, y = var_315_cast_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_65_cast_fp16 = add(x = var_2386_cast_fp16, y = var_2393_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_46)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_48_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_48 = read_state(input = key_cache)[name = string("coreml_update_state_48")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2369_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_47)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_49_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_49 = read_state(input = value_cache)[name = string("coreml_update_state_49")];
            tensor<int32, [4]> var_2463_begin_0 = const()[name = string("op_2463_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2463_end_0 = const()[name = string("op_2463_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2463_end_mask_0 = const()[name = string("op_2463_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2463_cast_fp16 = slice_by_index(begin = var_2463_begin_0, end = var_2463_end_0, end_mask = var_2463_end_mask_0, x = coreml_update_state_48)[name = string("op_2463_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2466_axis_0 = const()[name = string("op_2466_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_1 = split(axis = var_2466_axis_0, split_sizes = tile_12, x = var_2463_cast_fp16)[name = string("op_2466_cast_fp16")];
            tensor<int32, [4]> var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, x = coreml_update_state_49)[name = string("op_2473_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2476_axis_0 = const()[name = string("op_2476_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_1 = split(axis = var_2476_axis_0, split_sizes = tile_13, x = var_2473_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<int32, [2]> var_2479_split_sizes_0 = const()[name = string("op_2479_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2479_axis_0 = const()[name = string("op_2479_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2479_0, tensor<fp16, [1, 8, 128, 8]> var_2479_1 = split(axis = var_2479_axis_0, split_sizes = var_2479_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2479")];
            bool attn_weights_97_transpose_x_0 = const()[name = string("attn_weights_97_transpose_x_0"), val = bool(false)];
            bool attn_weights_97_transpose_y_0 = const()[name = string("attn_weights_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_0, transpose_y = attn_weights_97_transpose_y_0, x = var_2466_cast_fp16_0, y = var_2479_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2482_to_fp16 = const()[name = string("op_2482_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_2482_to_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_101_cast_fp16 = add(x = attn_weights_99_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_101_cast_fp16")];
            int32 var_2486 = const()[name = string("op_2486"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_103_cast_fp16 = softmax(axis = var_2486, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2492_transpose_x_1 = const()[name = string("op_2492_transpose_x_1"), val = bool(true)];
            bool var_2492_transpose_y_1 = const()[name = string("op_2492_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2492_cast_fp16 = matmul(transpose_x = var_2492_transpose_x_1, transpose_y = var_2492_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2476_cast_fp16_0)[name = string("op_2492_cast_fp16")];
            bool attn_weights_105_transpose_x_0 = const()[name = string("attn_weights_105_transpose_x_0"), val = bool(false)];
            bool attn_weights_105_transpose_y_0 = const()[name = string("attn_weights_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_105_cast_fp16 = matmul(transpose_x = attn_weights_105_transpose_x_0, transpose_y = attn_weights_105_transpose_y_0, x = var_2466_cast_fp16_1, y = var_2479_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2494_to_fp16 = const()[name = string("op_2494_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_107_cast_fp16 = mul(x = attn_weights_105_cast_fp16, y = var_2494_to_fp16)[name = string("attn_weights_107_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_109_cast_fp16 = add(x = attn_weights_107_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_109_cast_fp16")];
            int32 var_2498 = const()[name = string("op_2498"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_111_cast_fp16 = softmax(axis = var_2498, x = attn_weights_109_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(true)];
            bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_111_cast_fp16, y = var_2476_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2506 = const()[name = string("op_2506"), val = int32(1)];
            bool attn_output_51_interleave_0 = const()[name = string("attn_output_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_51_cast_fp16 = concat(axis = var_2506, interleave = attn_output_51_interleave_0, values = (var_2492_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2510_perm_0 = const()[name = string("op_2510_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_2510_cast_fp16 = transpose(perm = var_2510_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2510_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2543_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2543_cast_fp16")];
            int32 var_2541 = const()[name = string("op_2541"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_53_cast_fp16 = concat(axis = var_2541, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2543_cast_fp16))[name = string("doubled_53_cast_fp16")];
            tensor<int32, [1]> out_27_axes_0 = const()[name = string("out_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_27_gamma_0_to_fp16 = const()[name = string("out_27_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381144640)))];
            fp16 var_2553_to_fp16 = const()[name = string("op_2553_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_2553_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2564_split_sizes_0 = const()[name = string("op_2564_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2564_axis_0 = const()[name = string("op_2564_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2564_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2564_cast_fp16_1 = split(axis = var_2564_axis_0, split_sizes = var_2564_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2581_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [2]> var_2587_strides_0 = const()[name = string("op_2587_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2587_pad_type_0 = const()[name = string("op_2587_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2587_pad_0 = const()[name = string("op_2587_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2587_dilations_0 = const()[name = string("op_2587_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2587_groups_0 = const()[name = string("op_2587_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2587_cast_fp16 = conv(dilations = var_2587_dilations_0, groups = var_2587_groups_0, pad = var_2587_pad_0, pad_type = var_2587_pad_type_0, strides = var_2587_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("op_2587_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_69_cast_fp16 = mul(x = var_2581_cast_fp16, y = var_2587_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381152896)))];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_to_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2605_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2605_cast_fp16")];
            int32 var_2603 = const()[name = string("op_2603"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_57_cast_fp16 = concat(axis = var_2603, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2605_cast_fp16))[name = string("doubled_57_cast_fp16")];
            tensor<int32, [1]> out_29_axes_0 = const()[name = string("out_29_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_29_gamma_0_to_fp16 = const()[name = string("out_29_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406318784)))];
            fp16 var_2615_to_fp16 = const()[name = string("op_2615_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_2615_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2626_split_sizes_0 = const()[name = string("op_2626_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2626_axis_0 = const()[name = string("op_2626_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2626_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2626_cast_fp16_1 = split(axis = var_2626_axis_0, split_sizes = var_2626_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2626_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406327040)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2626_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_2683_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_2690_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2690_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2694_cast_fp16 = mul(x = x_71_cast_fp16, y = var_308_cast_fp16)[name = string("op_2694_cast_fp16")];
            tensor<int32, [2]> var_2695_split_sizes_0 = const()[name = string("op_2695_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2695_axis_0 = const()[name = string("op_2695_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2695_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2695_cast_fp16_1 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2695_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_2697_cast_fp16 = mul(x = var_2695_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2697_cast_fp16")];
            int32 var_2699 = const()[name = string("op_2699"), val = int32(-2)];
            bool var_2700_interleave_0 = const()[name = string("op_2700_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2700_cast_fp16 = concat(axis = var_2699, interleave = var_2700_interleave_0, values = (var_2697_cast_fp16, var_2695_cast_fp16_0))[name = string("op_2700_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2701_cast_fp16 = mul(x = var_2700_cast_fp16, y = var_315_cast_fp16)[name = string("op_2701_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_45_cast_fp16 = add(x = var_2694_cast_fp16, y = var_2701_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2707_cast_fp16 = mul(x = var_2683_cast_fp16, y = var_308_cast_fp16)[name = string("op_2707_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2708_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = var_2683_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_2710_cast_fp16 = mul(x = var_2708_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2710_cast_fp16")];
            int32 var_2712 = const()[name = string("op_2712"), val = int32(-2)];
            bool var_2713_interleave_0 = const()[name = string("op_2713_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2713_cast_fp16 = concat(axis = var_2712, interleave = var_2713_interleave_0, values = (var_2710_cast_fp16, var_2708_cast_fp16_0))[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2714_cast_fp16 = mul(x = var_2713_cast_fp16, y = var_315_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_75_cast_fp16 = add(x = var_2707_cast_fp16, y = var_2714_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_48)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_50_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_50 = read_state(input = key_cache)[name = string("coreml_update_state_50")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2690_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_49)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_51_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_51 = read_state(input = value_cache)[name = string("coreml_update_state_51")];
            tensor<int32, [4]> var_2784_begin_0 = const()[name = string("op_2784_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2784_end_0 = const()[name = string("op_2784_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2784_end_mask_0 = const()[name = string("op_2784_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2784_cast_fp16 = slice_by_index(begin = var_2784_begin_0, end = var_2784_end_0, end_mask = var_2784_end_mask_0, x = coreml_update_state_50)[name = string("op_2784_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2787_axis_0 = const()[name = string("op_2787_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_1 = split(axis = var_2787_axis_0, split_sizes = tile_14, x = var_2784_cast_fp16)[name = string("op_2787_cast_fp16")];
            tensor<int32, [4]> var_2794_begin_0 = const()[name = string("op_2794_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2794_end_0 = const()[name = string("op_2794_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2794_end_mask_0 = const()[name = string("op_2794_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2794_cast_fp16 = slice_by_index(begin = var_2794_begin_0, end = var_2794_end_0, end_mask = var_2794_end_mask_0, x = coreml_update_state_51)[name = string("op_2794_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2797_axis_0 = const()[name = string("op_2797_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_1 = split(axis = var_2797_axis_0, split_sizes = tile_15, x = var_2794_cast_fp16)[name = string("op_2797_cast_fp16")];
            tensor<int32, [2]> var_2800_split_sizes_0 = const()[name = string("op_2800_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2800_axis_0 = const()[name = string("op_2800_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2800_0, tensor<fp16, [1, 8, 128, 8]> var_2800_1 = split(axis = var_2800_axis_0, split_sizes = var_2800_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2800")];
            bool attn_weights_113_transpose_x_0 = const()[name = string("attn_weights_113_transpose_x_0"), val = bool(false)];
            bool attn_weights_113_transpose_y_0 = const()[name = string("attn_weights_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_113_cast_fp16 = matmul(transpose_x = attn_weights_113_transpose_x_0, transpose_y = attn_weights_113_transpose_y_0, x = var_2787_cast_fp16_0, y = var_2800_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2803_to_fp16 = const()[name = string("op_2803_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_115_cast_fp16 = mul(x = attn_weights_113_cast_fp16, y = var_2803_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            int32 var_2807 = const()[name = string("op_2807"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_119_cast_fp16 = softmax(axis = var_2807, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2813_transpose_x_1 = const()[name = string("op_2813_transpose_x_1"), val = bool(true)];
            bool var_2813_transpose_y_1 = const()[name = string("op_2813_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2813_cast_fp16 = matmul(transpose_x = var_2813_transpose_x_1, transpose_y = var_2813_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2797_cast_fp16_0)[name = string("op_2813_cast_fp16")];
            bool attn_weights_121_transpose_x_0 = const()[name = string("attn_weights_121_transpose_x_0"), val = bool(false)];
            bool attn_weights_121_transpose_y_0 = const()[name = string("attn_weights_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_0, transpose_y = attn_weights_121_transpose_y_0, x = var_2787_cast_fp16_1, y = var_2800_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2815_to_fp16 = const()[name = string("op_2815_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_2815_to_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_125_cast_fp16 = add(x = attn_weights_123_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_125_cast_fp16")];
            int32 var_2819 = const()[name = string("op_2819"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_127_cast_fp16 = softmax(axis = var_2819, x = attn_weights_125_cast_fp16)[name = string("attn_weights_127_cast_fp16")];
            bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(true)];
            bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_127_cast_fp16, y = var_2797_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2827 = const()[name = string("op_2827"), val = int32(1)];
            bool attn_output_59_interleave_0 = const()[name = string("attn_output_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_59_cast_fp16 = concat(axis = var_2827, interleave = attn_output_59_interleave_0, values = (var_2813_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2831_perm_0 = const()[name = string("op_2831_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_2831_cast_fp16 = transpose(perm = var_2831_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2831_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2864_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2864_cast_fp16")];
            int32 var_2862 = const()[name = string("op_2862"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_61_cast_fp16 = concat(axis = var_2862, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2864_cast_fp16))[name = string("doubled_61_cast_fp16")];
            tensor<int32, [1]> out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_31_gamma_0_to_fp16 = const()[name = string("out_31_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407375680)))];
            fp16 var_2874_to_fp16 = const()[name = string("op_2874_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_2874_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2885_split_sizes_0 = const()[name = string("op_2885_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2885_axis_0 = const()[name = string("op_2885_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2885_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2885_cast_fp16_1 = split(axis = var_2885_axis_0, split_sizes = var_2885_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2885_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2885_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2902_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2902_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407383936)))];
            tensor<int32, [2]> var_2908_strides_0 = const()[name = string("op_2908_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2908_pad_type_0 = const()[name = string("op_2908_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2908_pad_0 = const()[name = string("op_2908_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2908_dilations_0 = const()[name = string("op_2908_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2908_groups_0 = const()[name = string("op_2908_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2908_cast_fp16 = conv(dilations = var_2908_dilations_0, groups = var_2908_groups_0, pad = var_2908_pad_0, pad_type = var_2908_pad_type_0, strides = var_2908_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2885_cast_fp16_0)[name = string("op_2908_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_79_cast_fp16 = mul(x = var_2902_cast_fp16, y = var_2908_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432549824)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2926_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2926_cast_fp16")];
            int32 var_2924 = const()[name = string("op_2924"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_65_cast_fp16 = concat(axis = var_2924, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2926_cast_fp16))[name = string("doubled_65_cast_fp16")];
            tensor<int32, [1]> out_33_axes_0 = const()[name = string("out_33_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_33_gamma_0_to_fp16 = const()[name = string("out_33_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457715712)))];
            fp16 var_2936_to_fp16 = const()[name = string("op_2936_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_2936_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2947_split_sizes_0 = const()[name = string("op_2947_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2947_axis_0 = const()[name = string("op_2947_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2947_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2947_cast_fp16_1 = split(axis = var_2947_axis_0, split_sizes = var_2947_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2947_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457723968)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466112640)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467161280)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_3004_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3004_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_3011_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3011_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3015_cast_fp16 = mul(x = x_81_cast_fp16, y = var_308_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<int32, [2]> var_3016_split_sizes_0 = const()[name = string("op_3016_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3016_axis_0 = const()[name = string("op_3016_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_3016_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_3016_cast_fp16_1 = split(axis = var_3016_axis_0, split_sizes = var_3016_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3016_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_3018_cast_fp16 = mul(x = var_3016_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3018_cast_fp16")];
            int32 var_3020 = const()[name = string("op_3020"), val = int32(-2)];
            bool var_3021_interleave_0 = const()[name = string("op_3021_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_3021_cast_fp16 = concat(axis = var_3020, interleave = var_3021_interleave_0, values = (var_3018_cast_fp16, var_3016_cast_fp16_0))[name = string("op_3021_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3022_cast_fp16 = mul(x = var_3021_cast_fp16, y = var_315_cast_fp16)[name = string("op_3022_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_51_cast_fp16 = add(x = var_3015_cast_fp16, y = var_3022_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3028_cast_fp16 = mul(x = var_3004_cast_fp16, y = var_308_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_3029_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = var_3004_cast_fp16)[name = string("op_3029_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_3031_cast_fp16 = mul(x = var_3029_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3031_cast_fp16")];
            int32 var_3033 = const()[name = string("op_3033"), val = int32(-2)];
            bool var_3034_interleave_0 = const()[name = string("op_3034_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_3034_cast_fp16 = concat(axis = var_3033, interleave = var_3034_interleave_0, values = (var_3031_cast_fp16, var_3029_cast_fp16_0))[name = string("op_3034_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3035_cast_fp16 = mul(x = var_3034_cast_fp16, y = var_315_cast_fp16)[name = string("op_3035_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_85_cast_fp16 = add(x = var_3028_cast_fp16, y = var_3035_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_50)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_52_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_52 = read_state(input = key_cache)[name = string("coreml_update_state_52")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3011_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_51)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_53_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_53 = read_state(input = value_cache)[name = string("coreml_update_state_53")];
            tensor<int32, [4]> var_3105_begin_0 = const()[name = string("op_3105_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3105_end_0 = const()[name = string("op_3105_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3105_end_mask_0 = const()[name = string("op_3105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3105_cast_fp16 = slice_by_index(begin = var_3105_begin_0, end = var_3105_end_0, end_mask = var_3105_end_mask_0, x = coreml_update_state_52)[name = string("op_3105_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3108_axis_0 = const()[name = string("op_3108_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_1 = split(axis = var_3108_axis_0, split_sizes = tile_16, x = var_3105_cast_fp16)[name = string("op_3108_cast_fp16")];
            tensor<int32, [4]> var_3115_begin_0 = const()[name = string("op_3115_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3115_end_0 = const()[name = string("op_3115_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3115_end_mask_0 = const()[name = string("op_3115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3115_cast_fp16 = slice_by_index(begin = var_3115_begin_0, end = var_3115_end_0, end_mask = var_3115_end_mask_0, x = coreml_update_state_53)[name = string("op_3115_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3118_axis_0 = const()[name = string("op_3118_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_1 = split(axis = var_3118_axis_0, split_sizes = tile_17, x = var_3115_cast_fp16)[name = string("op_3118_cast_fp16")];
            tensor<int32, [2]> var_3121_split_sizes_0 = const()[name = string("op_3121_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3121_axis_0 = const()[name = string("op_3121_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_3121_0, tensor<fp16, [1, 8, 128, 8]> var_3121_1 = split(axis = var_3121_axis_0, split_sizes = var_3121_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3121")];
            bool attn_weights_129_transpose_x_0 = const()[name = string("attn_weights_129_transpose_x_0"), val = bool(false)];
            bool attn_weights_129_transpose_y_0 = const()[name = string("attn_weights_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_129_cast_fp16 = matmul(transpose_x = attn_weights_129_transpose_x_0, transpose_y = attn_weights_129_transpose_y_0, x = var_3108_cast_fp16_0, y = var_3121_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3124_to_fp16 = const()[name = string("op_3124_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_131_cast_fp16 = mul(x = attn_weights_129_cast_fp16, y = var_3124_to_fp16)[name = string("attn_weights_131_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_133_cast_fp16 = add(x = attn_weights_131_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_133_cast_fp16")];
            int32 var_3128 = const()[name = string("op_3128"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_135_cast_fp16 = softmax(axis = var_3128, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3134_transpose_x_1 = const()[name = string("op_3134_transpose_x_1"), val = bool(true)];
            bool var_3134_transpose_y_1 = const()[name = string("op_3134_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_3134_cast_fp16 = matmul(transpose_x = var_3134_transpose_x_1, transpose_y = var_3134_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3118_cast_fp16_0)[name = string("op_3134_cast_fp16")];
            bool attn_weights_137_transpose_x_0 = const()[name = string("attn_weights_137_transpose_x_0"), val = bool(false)];
            bool attn_weights_137_transpose_y_0 = const()[name = string("attn_weights_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_137_cast_fp16 = matmul(transpose_x = attn_weights_137_transpose_x_0, transpose_y = attn_weights_137_transpose_y_0, x = var_3108_cast_fp16_1, y = var_3121_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3136_to_fp16 = const()[name = string("op_3136_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_139_cast_fp16 = mul(x = attn_weights_137_cast_fp16, y = var_3136_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            int32 var_3140 = const()[name = string("op_3140"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_cast_fp16 = softmax(axis = var_3140, x = attn_weights_141_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(true)];
            bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_cast_fp16, y = var_3118_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3148 = const()[name = string("op_3148"), val = int32(1)];
            bool attn_output_67_interleave_0 = const()[name = string("attn_output_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_67_cast_fp16 = concat(axis = var_3148, interleave = attn_output_67_interleave_0, values = (var_3134_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3152_perm_0 = const()[name = string("op_3152_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_3152_cast_fp16 = transpose(perm = var_3152_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_cast_fp16 = reshape(shape = concat_107x, x = var_3152_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468209920)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3185_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3185_cast_fp16")];
            int32 var_3183 = const()[name = string("op_3183"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_69_cast_fp16 = concat(axis = var_3183, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3185_cast_fp16))[name = string("doubled_69_cast_fp16")];
            tensor<int32, [1]> out_35_axes_0 = const()[name = string("out_35_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_35_gamma_0_to_fp16 = const()[name = string("out_35_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476598592)))];
            fp16 var_3195_to_fp16 = const()[name = string("op_3195_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_3195_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3206_split_sizes_0 = const()[name = string("op_3206_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3206_axis_0 = const()[name = string("op_3206_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3206_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3206_cast_fp16_1 = split(axis = var_3206_axis_0, split_sizes = var_3206_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3206_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476606848)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_8_mlp_gate_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_3223_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3223_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501772736)))];
            tensor<int32, [2]> var_3229_strides_0 = const()[name = string("op_3229_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3229_pad_type_0 = const()[name = string("op_3229_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3229_pad_0 = const()[name = string("op_3229_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3229_dilations_0 = const()[name = string("op_3229_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3229_groups_0 = const()[name = string("op_3229_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_3229_cast_fp16 = conv(dilations = var_3229_dilations_0, groups = var_3229_groups_0, pad = var_3229_pad_0, pad_type = var_3229_pad_type_0, strides = var_3229_strides_0, weight = layers_8_mlp_up_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("op_3229_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_cast_fp16 = mul(x = var_3223_cast_fp16, y = var_3229_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526938624)))];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3247_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3247_cast_fp16")];
            int32 var_3245 = const()[name = string("op_3245"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_73_cast_fp16 = concat(axis = var_3245, interleave = doubled_73_interleave_0, values = (hidden_states_cast_fp16, var_3247_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_axes_0 = const()[name = string("out_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_gamma_0_to_fp16 = const()[name = string("out_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552104512)))];
            fp16 var_3257_to_fp16 = const()[name = string("op_3257_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3257_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3268_split_sizes_0 = const()[name = string("op_3268_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3268_axis_0 = const()[name = string("op_3268_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states, tensor<fp16, [1, 2048, 1, 8]> var_3268_cast_fp16_1 = split(axis = var_3268_axis_0, split_sizes = var_3268_split_sizes_0, x = out_cast_fp16)[name = string("op_3268_cast_fp16")];
        } -> (hidden_states);
    func main<ios18>(tensor<fp32, [1, 2048, 1, 1]> inputs_embeds, state<tensor<fp16, [9, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [1]> position_index_seed, state<tensor<fp16, [9, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4194432))))[name = string("layers_0_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4200704))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725632))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725056))))[name = string("layers_0_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4725952))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17321280))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17308928))))[name = string("layers_0_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17327488))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29922816))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29910464))))[name = string("layers_0_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29929024))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42516160))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42512000))))[name = string("layers_0_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42518272))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46716800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46712640))))[name = string("layers_1_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46718912))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243840))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47243264))))[name = string("layers_1_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47244160))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51442688))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51438528))))[name = string("layers_1_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51444800))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64040128))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64027776))))[name = string("layers_1_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64046336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76633472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76629312))))[name = string("layers_1_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76635584))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80834112))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80829952))))[name = string("layers_2_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80836224))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361152))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81360576))))[name = string("layers_2_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81361472))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85560000))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85555840))))[name = string("layers_2_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85562112))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98157440))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98145088))))[name = string("layers_2_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98163648))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110758976))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110746624))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110765184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123352320))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123348160))))[name = string("layers_2_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123354432))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127552960))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127548800))))[name = string("layers_3_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127555072))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080000))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128079424))))[name = string("layers_3_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128080320))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140675648))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140663296))))[name = string("layers_3_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140681856))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153277184))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153264832))))[name = string("layers_3_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_3_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153283392))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165870528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165866368))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165872640))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170071168))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170067008))))[name = string("layers_4_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170073280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170597632))))[name = string("layers_4_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170598528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174797056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174792896))))[name = string("layers_4_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174799168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187394496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187382144))))[name = string("layers_4_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_4_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187400704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199996032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199983680))))[name = string("layers_4_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_4_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200002240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212589376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212585216))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212591488))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216790016))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216785856))))[name = string("layers_5_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216792128))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317056))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217316480))))[name = string("layers_5_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217317376))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221515904))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221511744))))[name = string("layers_5_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221518016))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234113344))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234100992))))[name = string("layers_5_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234119552))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246714880))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246702528))))[name = string("layers_5_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_5_mlp_down_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246721088))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259308224))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259304064))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259310336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263508864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263504704))))[name = string("layers_6_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263510976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264035328))))[name = string("layers_6_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264036224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268234752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268230592))))[name = string("layers_6_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268236864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280832192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280819840))))[name = string("layers_6_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_6_mlp_up_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280838400))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293433728))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293421376))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293439936))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297638464))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297634304))))[name = string("layers_7_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297640576))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165504))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298164928))))[name = string("layers_7_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_o_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298165824))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302364352))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302360192))))[name = string("layers_7_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_gate_proj_weight_cast_fp16 = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302366464))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314961792))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314949440))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            int32 gather_0_cast_uint16_to_int32 = const()[name = string("gather_0_cast_uint16_to_int32"), val = int32(1)];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = gather_0_cast_uint16_to_int32)[name = string("cache_position_end")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> fill_like_0_cast_fp16 = fill_like(ref_tensor = position_index_seed, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            int32 var_280 = const()[name = string("op_280"), val = int32(0)];
            bool var_282_exclusive_0 = const()[name = string("op_282_exclusive_0"), val = bool(false)];
            bool var_282_reverse_0 = const()[name = string("op_282_reverse_0"), val = bool(false)];
            tensor<fp16, [1]> var_282_cast_fp16 = cumsum(axis = var_280, exclusive = var_282_exclusive_0, reverse = var_282_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_282_cast_fp16")];
            fp16 var_284_promoted_to_fp16 = const()[name = string("op_284_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> position_offsets_cast_fp16 = sub(x = var_282_cast_fp16, y = var_284_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1]> var_287_cast_fp16 = expand_dims(axes = var_287_axes_0, x = position_offsets_cast_fp16)[name = string("op_287_cast_fp16")];
            string position_id_promoted_to_fp16_dtype_0 = const()[name = string("position_id_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> position_id_to_fp16 = cast(dtype = position_id_promoted_to_fp16_dtype_0, x = position_id)[name = string("cast_3")];
            tensor<fp16, [1, 1]> position_ids_1_cast_fp16 = add(x = var_287_cast_fp16, y = position_id_to_fp16)[name = string("position_ids_1_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int32, [1, 1]> position_ids_1_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = position_ids_1_cast_fp16)[name = string("cast_2")];
            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = position_ids_1_cast_fp16_to_int32, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(32768)];
            tensor<int32, [1, 1]> add_0 = add(x = position_ids_1_cast_fp16_to_int32, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 1]> select_0 = select(a = position_ids_1_cast_fp16_to_int32, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            tensor<fp16, [32768, 128]> rope_emb_cos_cached_to_fp16 = const()[name = string("rope_emb_cos_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314968000)))];
            int32 cos_1_batch_dims_0 = const()[name = string("cos_1_batch_dims_0"), val = int32(0)];
            bool cos_1_validate_indices_0 = const()[name = string("cos_1_validate_indices_0"), val = bool(false)];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [1, 1]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(32768)];
            tensor<int32, [1, 1]> add_0_1 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0_1")];
            tensor<int32, [1, 1]> select_0_1 = select(a = select_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 cos_1_cast_fp16_axis_0 = const()[name = string("cos_1_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 1, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_0, batch_dims = cos_1_batch_dims_0, indices = select_0_1, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [32768, 128]> rope_emb_sin_cached_to_fp16 = const()[name = string("rope_emb_sin_cached_to_fp16"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323356672)))];
            int32 sin_1_batch_dims_0 = const()[name = string("sin_1_batch_dims_0"), val = int32(0)];
            bool sin_1_validate_indices_0 = const()[name = string("sin_1_validate_indices_0"), val = bool(false)];
            int32 sin_1_cast_fp16_axis_0 = const()[name = string("sin_1_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 1, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_0, batch_dims = sin_1_batch_dims_0, indices = select_0_1, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_306_perm_0 = const()[name = string("op_306_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_308_axes_0 = const()[name = string("op_308_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_306_cast_fp16 = transpose(perm = var_306_perm_0, x = cos_1_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 1, 128, 1]> var_308_cast_fp16 = expand_dims(axes = var_308_axes_0, x = var_306_cast_fp16)[name = string("op_308_cast_fp16")];
            tensor<int32, [3]> var_313_perm_0 = const()[name = string("op_313_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("op_315_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_313_cast_fp16 = transpose(perm = var_313_perm_0, x = sin_1_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 1, 128, 1]> var_315_cast_fp16 = expand_dims(axes = var_315_axes_0, x = var_313_cast_fp16)[name = string("op_315_cast_fp16")];
            tensor<int32, [1]> var_334_axes_0 = const()[name = string("op_334_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 1, 1]> var_334 = expand_dims(axes = var_334_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_334")];
            tensor<int32, [1, 1, 2048]> var_327 = const()[name = string("op_327"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331745344)))];
            tensor<bool, [1, 1, 2048]> var_335 = greater(x = var_327, y = var_334)[name = string("op_335")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_342_axes_0 = const()[name = string("op_342_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 2048]> var_335_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_335)[name = string("cast_1")];
            tensor<fp16, [1, 1, 1, 2048]> var_342_cast_fp16 = expand_dims(axes = var_342_axes_0, x = var_335_to_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_346_promoted_to_fp16 = const()[name = string("op_346_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 1]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_342_cast_fp16)[name = string("transpose_27")];
            tensor<bool, [1, 1, 2048, 1]> var_347_cast_fp16 = equal(x = mask_cast_fp16, y = var_346_promoted_to_fp16)[name = string("op_347_cast_fp16")];
            fp16 var_348_to_fp16 = const()[name = string("op_348_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 1]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_348_to_fp16, cond = var_347_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> inputs_embeds_to_fp16 = cast(dtype = inputs_embeds_to_fp16_dtype_0, x = inputs_embeds)[name = string("cast_0")];
            tensor<fp16, [1, 2048, 1, 1]> var_358_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_358_cast_fp16")];
            int32 var_356 = const()[name = string("op_356"), val = int32(1)];
            bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_1_cast_fp16 = concat(axis = var_356, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_358_cast_fp16))[name = string("doubled_1_cast_fp16")];
            tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_1_gamma_0_to_fp16 = const()[name = string("out_1_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331753600)))];
            fp16 var_368_to_fp16 = const()[name = string("op_368_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_368_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_379_split_sizes_0 = const()[name = string("op_379_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_379_axis_0 = const()[name = string("op_379_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_379_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_379_cast_fp16_1 = split(axis = var_379_axis_0, split_sizes = var_379_split_sizes_0, x = out_1_cast_fp16)[name = string("op_379_cast_fp16")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = layers_0_self_attn_q_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = layers_0_self_attn_k_proj_weight_cast_fp16, x = var_379_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331761856)))];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = var_379_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_1_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_436_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_436_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_443_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_443_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_447_cast_fp16 = mul(x = x_1_cast_fp16, y = var_308_cast_fp16)[name = string("op_447_cast_fp16")];
            tensor<int32, [2]> var_448_split_sizes_0 = const()[name = string("op_448_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_448_axis_0 = const()[name = string("op_448_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_448_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_448_cast_fp16_1 = split(axis = var_448_axis_0, split_sizes = var_448_split_sizes_0, x = x_1_cast_fp16)[name = string("op_448_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_450_cast_fp16 = mul(x = var_448_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_450_cast_fp16")];
            int32 var_452 = const()[name = string("op_452"), val = int32(-2)];
            bool var_453_interleave_0 = const()[name = string("op_453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_453_cast_fp16 = concat(axis = var_452, interleave = var_453_interleave_0, values = (var_450_cast_fp16, var_448_cast_fp16_0))[name = string("op_453_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_454_cast_fp16 = mul(x = var_453_cast_fp16, y = var_315_cast_fp16)[name = string("op_454_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_3_cast_fp16 = add(x = var_447_cast_fp16, y = var_454_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_460_cast_fp16 = mul(x = var_436_cast_fp16, y = var_308_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [2]> var_461_split_sizes_0 = const()[name = string("op_461_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_461_axis_0 = const()[name = string("op_461_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_461_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_461_cast_fp16_1 = split(axis = var_461_axis_0, split_sizes = var_461_split_sizes_0, x = var_436_cast_fp16)[name = string("op_461_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_463_cast_fp16 = mul(x = var_461_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_463_cast_fp16")];
            int32 var_465 = const()[name = string("op_465"), val = int32(-2)];
            bool var_466_interleave_0 = const()[name = string("op_466_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_466_cast_fp16 = concat(axis = var_465, interleave = var_466_interleave_0, values = (var_463_cast_fp16, var_461_cast_fp16_0))[name = string("op_466_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_467_cast_fp16 = mul(x = var_466_cast_fp16, y = var_315_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_5_cast_fp16 = add(x = var_460_cast_fp16, y = var_467_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (expand_dims_0, expand_dims_1, position_id, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (expand_dims_4, concat_6_values1_0, cache_position_end, concat_6_values3_0))[name = string("concat_6")];
            tensor<int32, [4]> key_states_7_perm_0 = const()[name = string("key_states_7_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_7_cast_fp16 = transpose(perm = key_states_7_perm_0, x = key_states_5_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = key_states_7_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_0_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_0 = read_state(input = key_cache)[name = string("coreml_update_state_0")];
            tensor<fp16, [9, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = string("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_443_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = value_states_3_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_1_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_1 = read_state(input = value_cache)[name = string("coreml_update_state_1")];
            tensor<int32, [4]> var_537_begin_0 = const()[name = string("op_537_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_537_end_0 = const()[name = string("op_537_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_537_end_mask_0 = const()[name = string("op_537_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_537_cast_fp16 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = coreml_update_state_0)[name = string("op_537_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_540_axis_0 = const()[name = string("op_540_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_540_cast_fp16_1 = split(axis = var_540_axis_0, split_sizes = tile_0, x = var_537_cast_fp16)[name = string("op_540_cast_fp16")];
            tensor<int32, [4]> var_547_begin_0 = const()[name = string("op_547_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_547_end_0 = const()[name = string("op_547_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_547_end_mask_0 = const()[name = string("op_547_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_547_cast_fp16 = slice_by_index(begin = var_547_begin_0, end = var_547_end_0, end_mask = var_547_end_mask_0, x = coreml_update_state_1)[name = string("op_547_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_550_axis_0 = const()[name = string("op_550_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_550_cast_fp16_1 = split(axis = var_550_axis_0, split_sizes = tile_1, x = var_547_cast_fp16)[name = string("op_550_cast_fp16")];
            tensor<int32, [2]> var_553_split_sizes_0 = const()[name = string("op_553_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_553_axis_0 = const()[name = string("op_553_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_553_0, tensor<fp16, [1, 8, 128, 1]> var_553_1 = split(axis = var_553_axis_0, split_sizes = var_553_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_553")];
            bool attn_weights_1_transpose_x_0 = const()[name = string("attn_weights_1_transpose_x_0"), val = bool(false)];
            bool attn_weights_1_transpose_y_0 = const()[name = string("attn_weights_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_540_cast_fp16_0, y = var_553_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_556_to_fp16 = const()[name = string("op_556_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_556_to_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_5_cast_fp16 = add(x = attn_weights_3_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_5_cast_fp16")];
            int32 var_560 = const()[name = string("op_560"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_7_cast_fp16 = softmax(axis = var_560, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_566_transpose_x_1 = const()[name = string("op_566_transpose_x_1"), val = bool(true)];
            bool var_566_transpose_y_1 = const()[name = string("op_566_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_566_cast_fp16 = matmul(transpose_x = var_566_transpose_x_1, transpose_y = var_566_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_550_cast_fp16_0)[name = string("op_566_cast_fp16")];
            bool attn_weights_9_transpose_x_0 = const()[name = string("attn_weights_9_transpose_x_0"), val = bool(false)];
            bool attn_weights_9_transpose_y_0 = const()[name = string("attn_weights_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_9_cast_fp16 = matmul(transpose_x = attn_weights_9_transpose_x_0, transpose_y = attn_weights_9_transpose_y_0, x = var_540_cast_fp16_1, y = var_553_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_568_to_fp16 = const()[name = string("op_568_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_11_cast_fp16 = mul(x = attn_weights_9_cast_fp16, y = var_568_to_fp16)[name = string("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_13_cast_fp16 = add(x = attn_weights_11_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_13_cast_fp16")];
            int32 var_572 = const()[name = string("op_572"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_15_cast_fp16 = softmax(axis = var_572, x = attn_weights_13_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            bool attn_output_1_transpose_x_1 = const()[name = string("attn_output_1_transpose_x_1"), val = bool(true)];
            bool attn_output_1_transpose_y_1 = const()[name = string("attn_output_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_1, transpose_y = attn_output_1_transpose_y_1, x = attn_weights_15_cast_fp16, y = var_550_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_580 = const()[name = string("op_580"), val = int32(1)];
            bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_3_cast_fp16 = concat(axis = var_580, interleave = attn_output_3_interleave_0, values = (var_566_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_584_perm_0 = const()[name = string("op_584_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_584_cast_fp16 = transpose(perm = var_584_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_584_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332810496)))];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_3_cast_fp16 = conv(dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = attn_output_7_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_5_cast_fp16 = add(x = inputs_embeds_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_617_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_617_cast_fp16")];
            int32 var_615 = const()[name = string("op_615"), val = int32(1)];
            bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_5_cast_fp16 = concat(axis = var_615, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_617_cast_fp16))[name = string("doubled_5_cast_fp16")];
            tensor<int32, [1]> out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_3_gamma_0_to_fp16 = const()[name = string("out_3_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341199168)))];
            fp16 var_627_to_fp16 = const()[name = string("op_627_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_627_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_638_split_sizes_0 = const()[name = string("op_638_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_638_axis_0 = const()[name = string("op_638_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_638_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_638_cast_fp16_1 = split(axis = var_638_axis_0, split_sizes = var_638_split_sizes_0, x = out_3_cast_fp16)[name = string("op_638_cast_fp16")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_mlp_gate_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_655_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_655_cast_fp16")];
            tensor<int32, [2]> var_661_strides_0 = const()[name = string("op_661_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_661_pad_type_0 = const()[name = string("op_661_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_661_pad_0 = const()[name = string("op_661_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_661_dilations_0 = const()[name = string("op_661_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_661_groups_0 = const()[name = string("op_661_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_661_cast_fp16 = conv(dilations = var_661_dilations_0, groups = var_661_groups_0, pad = var_661_pad_0, pad_type = var_661_pad_type_0, strides = var_661_strides_0, weight = layers_0_mlp_up_proj_weight_cast_fp16, x = var_638_cast_fp16_0)[name = string("op_661_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_9_cast_fp16 = mul(x = var_655_cast_fp16, y = var_661_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_7_cast_fp16 = conv(dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_0_mlp_down_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_9_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_679_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_679_cast_fp16")];
            int32 var_677 = const()[name = string("op_677"), val = int32(1)];
            bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_9_cast_fp16 = concat(axis = var_677, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_679_cast_fp16))[name = string("doubled_9_cast_fp16")];
            tensor<int32, [1]> out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_5_gamma_0_to_fp16 = const()[name = string("out_5_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341207424)))];
            fp16 var_689_to_fp16 = const()[name = string("op_689_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_689_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_700_split_sizes_0 = const()[name = string("op_700_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_700_axis_0 = const()[name = string("op_700_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_700_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_700_cast_fp16_1 = split(axis = var_700_axis_0, split_sizes = var_700_split_sizes_0, x = out_5_cast_fp16)[name = string("op_700_cast_fp16")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_7_pad_0 = const()[name = string("query_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_7_dilations_0 = const()[name = string("query_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_7_groups_0 = const()[name = string("query_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_7_cast_fp16 = conv(dilations = query_states_7_dilations_0, groups = query_states_7_groups_0, pad = query_states_7_pad_0, pad_type = query_states_7_pad_type_0, strides = query_states_7_strides_0, weight = layers_1_self_attn_q_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_11_pad_0 = const()[name = string("key_states_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_11_dilations_0 = const()[name = string("key_states_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_11_groups_0 = const()[name = string("key_states_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_11_cast_fp16 = conv(dilations = key_states_11_dilations_0, groups = key_states_11_groups_0, pad = key_states_11_pad_0, pad_type = key_states_11_pad_type_0, strides = key_states_11_strides_0, weight = layers_1_self_attn_k_proj_weight_cast_fp16, x = var_700_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341215680)))];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_7_pad_0 = const()[name = string("value_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_7_dilations_0 = const()[name = string("value_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_7_groups_0 = const()[name = string("value_states_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_7_cast_fp16 = conv(dilations = value_states_7_dilations_0, groups = value_states_7_groups_0, pad = value_states_7_pad_0, pad_type = value_states_7_pad_type_0, strides = value_states_7_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = var_700_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_11_cast_fp16 = reshape(shape = concat_12x, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_757_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_757_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_764_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_764_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_768_cast_fp16 = mul(x = x_11_cast_fp16, y = var_308_cast_fp16)[name = string("op_768_cast_fp16")];
            tensor<int32, [2]> var_769_split_sizes_0 = const()[name = string("op_769_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_769_axis_0 = const()[name = string("op_769_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_769_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_769_cast_fp16_1 = split(axis = var_769_axis_0, split_sizes = var_769_split_sizes_0, x = x_11_cast_fp16)[name = string("op_769_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_771_cast_fp16 = mul(x = var_769_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_771_cast_fp16")];
            int32 var_773 = const()[name = string("op_773"), val = int32(-2)];
            bool var_774_interleave_0 = const()[name = string("op_774_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_774_cast_fp16 = concat(axis = var_773, interleave = var_774_interleave_0, values = (var_771_cast_fp16, var_769_cast_fp16_0))[name = string("op_774_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_775_cast_fp16 = mul(x = var_774_cast_fp16, y = var_315_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_9_cast_fp16 = add(x = var_768_cast_fp16, y = var_775_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_781_cast_fp16 = mul(x = var_757_cast_fp16, y = var_308_cast_fp16)[name = string("op_781_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_782_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = var_757_cast_fp16)[name = string("op_782_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_784_cast_fp16 = mul(x = var_782_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_784_cast_fp16")];
            int32 var_786 = const()[name = string("op_786"), val = int32(-2)];
            bool var_787_interleave_0 = const()[name = string("op_787_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_787_cast_fp16 = concat(axis = var_786, interleave = var_787_interleave_0, values = (var_784_cast_fp16, var_782_cast_fp16_0))[name = string("op_787_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_788_cast_fp16 = mul(x = var_787_cast_fp16, y = var_315_cast_fp16)[name = string("op_788_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_15_cast_fp16 = add(x = var_781_cast_fp16, y = var_788_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (expand_dims_12, expand_dims_13, position_id, expand_dims_15))[name = string("concat_17")];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (expand_dims_16, concat_18_values1_0, cache_position_end, concat_18_values3_0))[name = string("concat_18")];
            tensor<int32, [4]> key_states_17_perm_0 = const()[name = string("key_states_17_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_17_cast_fp16 = transpose(perm = key_states_17_perm_0, x = key_states_15_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = key_states_17_cast_fp16, x = coreml_update_state_0)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_2_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_2 = read_state(input = key_cache)[name = string("coreml_update_state_2")];
            tensor<int32, [4]> value_states_9_perm_0 = const()[name = string("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_9_cast_fp16 = transpose(perm = value_states_9_perm_0, x = var_764_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_17, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_18, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = value_states_9_cast_fp16, x = coreml_update_state_1)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_3_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_3 = read_state(input = value_cache)[name = string("coreml_update_state_3")];
            tensor<int32, [4]> var_858_begin_0 = const()[name = string("op_858_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_858_end_0 = const()[name = string("op_858_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_858_end_mask_0 = const()[name = string("op_858_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_858_cast_fp16 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = coreml_update_state_2)[name = string("op_858_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_861_axis_0 = const()[name = string("op_861_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_861_cast_fp16_1 = split(axis = var_861_axis_0, split_sizes = tile_2, x = var_858_cast_fp16)[name = string("op_861_cast_fp16")];
            tensor<int32, [4]> var_868_begin_0 = const()[name = string("op_868_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_868_end_0 = const()[name = string("op_868_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_868_end_mask_0 = const()[name = string("op_868_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_868_cast_fp16 = slice_by_index(begin = var_868_begin_0, end = var_868_end_0, end_mask = var_868_end_mask_0, x = coreml_update_state_3)[name = string("op_868_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_871_axis_0 = const()[name = string("op_871_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_871_cast_fp16_1 = split(axis = var_871_axis_0, split_sizes = tile_3, x = var_868_cast_fp16)[name = string("op_871_cast_fp16")];
            tensor<int32, [2]> var_874_split_sizes_0 = const()[name = string("op_874_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_874_axis_0 = const()[name = string("op_874_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_874_0, tensor<fp16, [1, 8, 128, 1]> var_874_1 = split(axis = var_874_axis_0, split_sizes = var_874_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_874")];
            bool attn_weights_17_transpose_x_0 = const()[name = string("attn_weights_17_transpose_x_0"), val = bool(false)];
            bool attn_weights_17_transpose_y_0 = const()[name = string("attn_weights_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_17_cast_fp16 = matmul(transpose_x = attn_weights_17_transpose_x_0, transpose_y = attn_weights_17_transpose_y_0, x = var_861_cast_fp16_0, y = var_874_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_877_to_fp16 = const()[name = string("op_877_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_19_cast_fp16 = mul(x = attn_weights_17_cast_fp16, y = var_877_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            int32 var_881 = const()[name = string("op_881"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_23_cast_fp16 = softmax(axis = var_881, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_887_transpose_x_1 = const()[name = string("op_887_transpose_x_1"), val = bool(true)];
            bool var_887_transpose_y_1 = const()[name = string("op_887_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_887_cast_fp16 = matmul(transpose_x = var_887_transpose_x_1, transpose_y = var_887_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_871_cast_fp16_0)[name = string("op_887_cast_fp16")];
            bool attn_weights_25_transpose_x_0 = const()[name = string("attn_weights_25_transpose_x_0"), val = bool(false)];
            bool attn_weights_25_transpose_y_0 = const()[name = string("attn_weights_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_0, transpose_y = attn_weights_25_transpose_y_0, x = var_861_cast_fp16_1, y = var_874_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_889_to_fp16 = const()[name = string("op_889_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_889_to_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_29_cast_fp16 = add(x = attn_weights_27_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_29_cast_fp16")];
            int32 var_893 = const()[name = string("op_893"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_31_cast_fp16 = softmax(axis = var_893, x = attn_weights_29_cast_fp16)[name = string("attn_weights_31_cast_fp16")];
            bool attn_output_9_transpose_x_1 = const()[name = string("attn_output_9_transpose_x_1"), val = bool(true)];
            bool attn_output_9_transpose_y_1 = const()[name = string("attn_output_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_1, transpose_y = attn_output_9_transpose_y_1, x = attn_weights_31_cast_fp16, y = var_871_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_901 = const()[name = string("op_901"), val = int32(1)];
            bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_11_cast_fp16 = concat(axis = var_901, interleave = attn_output_11_interleave_0, values = (var_887_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_905_perm_0 = const()[name = string("op_905_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_905_cast_fp16 = transpose(perm = var_905_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_905_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_13_cast_fp16 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_1_self_attn_o_proj_weight_cast_fp16, x = attn_output_15_cast_fp16)[name = string("hidden_states_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_15_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_938_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_938_cast_fp16")];
            int32 var_936 = const()[name = string("op_936"), val = int32(1)];
            bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_13_cast_fp16 = concat(axis = var_936, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_938_cast_fp16))[name = string("doubled_13_cast_fp16")];
            tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_7_gamma_0_to_fp16 = const()[name = string("out_7_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342264320)))];
            fp16 var_948_to_fp16 = const()[name = string("op_948_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_948_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_959_split_sizes_0 = const()[name = string("op_959_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_959_axis_0 = const()[name = string("op_959_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_959_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_959_cast_fp16_1 = split(axis = var_959_axis_0, split_sizes = var_959_split_sizes_0, x = out_7_cast_fp16)[name = string("op_959_cast_fp16")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_1_mlp_gate_proj_weight_cast_fp16, x = var_959_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_976_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_976_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342272576)))];
            tensor<int32, [2]> var_982_strides_0 = const()[name = string("op_982_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_982_pad_type_0 = const()[name = string("op_982_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_982_pad_0 = const()[name = string("op_982_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_982_dilations_0 = const()[name = string("op_982_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_982_groups_0 = const()[name = string("op_982_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_982_cast_fp16 = conv(dilations = var_982_dilations_0, groups = var_982_groups_0, pad = var_982_pad_0, pad_type = var_982_pad_type_0, strides = var_982_strides_0, weight = layers_1_mlp_up_proj_weight_to_fp16, x = var_959_cast_fp16_0)[name = string("op_982_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_19_cast_fp16 = mul(x = var_976_cast_fp16, y = var_982_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_17_cast_fp16 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_1_mlp_down_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_19_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1000_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1000_cast_fp16")];
            int32 var_998 = const()[name = string("op_998"), val = int32(1)];
            bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_17_cast_fp16 = concat(axis = var_998, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1000_cast_fp16))[name = string("doubled_17_cast_fp16")];
            tensor<int32, [1]> out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_9_gamma_0_to_fp16 = const()[name = string("out_9_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367438464)))];
            fp16 var_1010_to_fp16 = const()[name = string("op_1010_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1010_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1021_split_sizes_0 = const()[name = string("op_1021_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1021_axis_0 = const()[name = string("op_1021_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1021_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1021_cast_fp16_1 = split(axis = var_1021_axis_0, split_sizes = var_1021_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1021_cast_fp16")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = layers_2_self_attn_q_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = layers_2_self_attn_k_proj_weight_cast_fp16, x = var_1021_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367446720)))];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = var_1021_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_21_cast_fp16 = reshape(shape = concat_24x, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1078_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1078_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1085_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1085_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1089_cast_fp16 = mul(x = x_21_cast_fp16, y = var_308_cast_fp16)[name = string("op_1089_cast_fp16")];
            tensor<int32, [2]> var_1090_split_sizes_0 = const()[name = string("op_1090_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1090_axis_0 = const()[name = string("op_1090_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1090_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1090_cast_fp16_1 = split(axis = var_1090_axis_0, split_sizes = var_1090_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1090_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_1092_cast_fp16 = mul(x = var_1090_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            int32 var_1094 = const()[name = string("op_1094"), val = int32(-2)];
            bool var_1095_interleave_0 = const()[name = string("op_1095_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1095_cast_fp16 = concat(axis = var_1094, interleave = var_1095_interleave_0, values = (var_1092_cast_fp16, var_1090_cast_fp16_0))[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1096_cast_fp16 = mul(x = var_1095_cast_fp16, y = var_315_cast_fp16)[name = string("op_1096_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_15_cast_fp16 = add(x = var_1089_cast_fp16, y = var_1096_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1102_cast_fp16 = mul(x = var_1078_cast_fp16, y = var_308_cast_fp16)[name = string("op_1102_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1103_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = var_1078_cast_fp16)[name = string("op_1103_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_1105_cast_fp16 = mul(x = var_1103_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1105_cast_fp16")];
            int32 var_1107 = const()[name = string("op_1107"), val = int32(-2)];
            bool var_1108_interleave_0 = const()[name = string("op_1108_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1108_cast_fp16 = concat(axis = var_1107, interleave = var_1108_interleave_0, values = (var_1105_cast_fp16, var_1103_cast_fp16_0))[name = string("op_1108_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1109_cast_fp16 = mul(x = var_1108_cast_fp16, y = var_315_cast_fp16)[name = string("op_1109_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_25_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1109_cast_fp16)[name = string("key_states_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (expand_dims_24, expand_dims_25, position_id, expand_dims_27))[name = string("concat_29")];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (expand_dims_28, concat_30_values1_0, cache_position_end, concat_30_values3_0))[name = string("concat_30")];
            tensor<int32, [4]> key_states_27_perm_0 = const()[name = string("key_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_27_cast_fp16 = transpose(perm = key_states_27_perm_0, x = key_states_25_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = key_states_27_cast_fp16, x = coreml_update_state_2)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_4_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_4 = read_state(input = key_cache)[name = string("coreml_update_state_4")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = string("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_1085_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_29, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_30, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = value_states_15_cast_fp16, x = coreml_update_state_3)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_5_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_5 = read_state(input = value_cache)[name = string("coreml_update_state_5")];
            tensor<int32, [4]> var_1179_begin_0 = const()[name = string("op_1179_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1179_end_0 = const()[name = string("op_1179_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1179_end_mask_0 = const()[name = string("op_1179_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1179_cast_fp16 = slice_by_index(begin = var_1179_begin_0, end = var_1179_end_0, end_mask = var_1179_end_mask_0, x = coreml_update_state_4)[name = string("op_1179_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1182_axis_0 = const()[name = string("op_1182_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1182_cast_fp16_1 = split(axis = var_1182_axis_0, split_sizes = tile_4, x = var_1179_cast_fp16)[name = string("op_1182_cast_fp16")];
            tensor<int32, [4]> var_1189_begin_0 = const()[name = string("op_1189_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1189_end_0 = const()[name = string("op_1189_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1189_end_mask_0 = const()[name = string("op_1189_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1189_cast_fp16 = slice_by_index(begin = var_1189_begin_0, end = var_1189_end_0, end_mask = var_1189_end_mask_0, x = coreml_update_state_5)[name = string("op_1189_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1192_axis_0 = const()[name = string("op_1192_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1192_cast_fp16_1 = split(axis = var_1192_axis_0, split_sizes = tile_5, x = var_1189_cast_fp16)[name = string("op_1192_cast_fp16")];
            tensor<int32, [2]> var_1195_split_sizes_0 = const()[name = string("op_1195_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1195_axis_0 = const()[name = string("op_1195_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1195_0, tensor<fp16, [1, 8, 128, 1]> var_1195_1 = split(axis = var_1195_axis_0, split_sizes = var_1195_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1195")];
            bool attn_weights_33_transpose_x_0 = const()[name = string("attn_weights_33_transpose_x_0"), val = bool(false)];
            bool attn_weights_33_transpose_y_0 = const()[name = string("attn_weights_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_33_cast_fp16 = matmul(transpose_x = attn_weights_33_transpose_x_0, transpose_y = attn_weights_33_transpose_y_0, x = var_1182_cast_fp16_0, y = var_1195_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1198_to_fp16 = const()[name = string("op_1198_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_35_cast_fp16 = mul(x = attn_weights_33_cast_fp16, y = var_1198_to_fp16)[name = string("attn_weights_35_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_37_cast_fp16 = add(x = attn_weights_35_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_37_cast_fp16")];
            int32 var_1202 = const()[name = string("op_1202"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_39_cast_fp16 = softmax(axis = var_1202, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1208_transpose_x_1 = const()[name = string("op_1208_transpose_x_1"), val = bool(true)];
            bool var_1208_transpose_y_1 = const()[name = string("op_1208_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1208_cast_fp16 = matmul(transpose_x = var_1208_transpose_x_1, transpose_y = var_1208_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1192_cast_fp16_0)[name = string("op_1208_cast_fp16")];
            bool attn_weights_41_transpose_x_0 = const()[name = string("attn_weights_41_transpose_x_0"), val = bool(false)];
            bool attn_weights_41_transpose_y_0 = const()[name = string("attn_weights_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_41_cast_fp16 = matmul(transpose_x = attn_weights_41_transpose_x_0, transpose_y = attn_weights_41_transpose_y_0, x = var_1182_cast_fp16_1, y = var_1195_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1210_to_fp16 = const()[name = string("op_1210_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_43_cast_fp16 = mul(x = attn_weights_41_cast_fp16, y = var_1210_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            int32 var_1214 = const()[name = string("op_1214"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_47_cast_fp16 = softmax(axis = var_1214, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")];
            bool attn_output_17_transpose_x_1 = const()[name = string("attn_output_17_transpose_x_1"), val = bool(true)];
            bool attn_output_17_transpose_y_1 = const()[name = string("attn_output_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_1, transpose_y = attn_output_17_transpose_y_1, x = attn_weights_47_cast_fp16, y = var_1192_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1222 = const()[name = string("op_1222"), val = int32(1)];
            bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_19_cast_fp16 = concat(axis = var_1222, interleave = attn_output_19_interleave_0, values = (var_1208_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1226_perm_0 = const()[name = string("op_1226_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_1226_cast_fp16 = transpose(perm = var_1226_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1226_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_23_cast_fp16 = conv(dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_2_self_attn_o_proj_weight_cast_fp16, x = attn_output_23_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_25_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("hidden_states_25_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1259_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            int32 var_1257 = const()[name = string("op_1257"), val = int32(1)];
            bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_21_cast_fp16 = concat(axis = var_1257, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1259_cast_fp16))[name = string("doubled_21_cast_fp16")];
            tensor<int32, [1]> out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_11_gamma_0_to_fp16 = const()[name = string("out_11_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368495360)))];
            fp16 var_1269_to_fp16 = const()[name = string("op_1269_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1269_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1280_split_sizes_0 = const()[name = string("op_1280_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1280_axis_0 = const()[name = string("op_1280_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1280_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1280_cast_fp16_1 = split(axis = var_1280_axis_0, split_sizes = var_1280_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1280_cast_fp16")];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_2_mlp_gate_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1297_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1297_cast_fp16")];
            tensor<int32, [2]> var_1303_strides_0 = const()[name = string("op_1303_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1303_pad_type_0 = const()[name = string("op_1303_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1303_pad_0 = const()[name = string("op_1303_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1303_dilations_0 = const()[name = string("op_1303_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1303_groups_0 = const()[name = string("op_1303_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1303_cast_fp16 = conv(dilations = var_1303_dilations_0, groups = var_1303_groups_0, pad = var_1303_pad_0, pad_type = var_1303_pad_type_0, strides = var_1303_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1280_cast_fp16_0)[name = string("op_1303_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_29_cast_fp16 = mul(x = var_1297_cast_fp16, y = var_1303_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_27_cast_fp16 = conv(dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_2_mlp_down_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_29_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1321_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1321_cast_fp16")];
            int32 var_1319 = const()[name = string("op_1319"), val = int32(1)];
            bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_25_cast_fp16 = concat(axis = var_1319, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1321_cast_fp16))[name = string("doubled_25_cast_fp16")];
            tensor<int32, [1]> out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_13_gamma_0_to_fp16 = const()[name = string("out_13_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368503616)))];
            fp16 var_1331_to_fp16 = const()[name = string("op_1331_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1331_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1342_split_sizes_0 = const()[name = string("op_1342_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1342_axis_0 = const()[name = string("op_1342_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1342_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1342_cast_fp16_1 = split(axis = var_1342_axis_0, split_sizes = var_1342_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1342_cast_fp16")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_19_pad_0 = const()[name = string("query_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_19_dilations_0 = const()[name = string("query_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_19_groups_0 = const()[name = string("query_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_19_cast_fp16 = conv(dilations = query_states_19_dilations_0, groups = query_states_19_groups_0, pad = query_states_19_pad_0, pad_type = query_states_19_pad_type_0, strides = query_states_19_strides_0, weight = layers_3_self_attn_q_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_31_pad_0 = const()[name = string("key_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_31_dilations_0 = const()[name = string("key_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_31_groups_0 = const()[name = string("key_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_31_cast_fp16 = conv(dilations = key_states_31_dilations_0, groups = key_states_31_groups_0, pad = key_states_31_pad_0, pad_type = key_states_31_pad_type_0, strides = key_states_31_strides_0, weight = layers_3_self_attn_k_proj_weight_cast_fp16, x = var_1342_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368511872)))];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_19_pad_0 = const()[name = string("value_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_19_dilations_0 = const()[name = string("value_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_19_groups_0 = const()[name = string("value_states_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_19_cast_fp16 = conv(dilations = value_states_19_dilations_0, groups = value_states_19_groups_0, pad = value_states_19_pad_0, pad_type = value_states_19_pad_type_0, strides = value_states_19_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = var_1342_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_31_cast_fp16 = reshape(shape = concat_36x, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1399_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1406_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1406_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1410_cast_fp16 = mul(x = x_31_cast_fp16, y = var_308_cast_fp16)[name = string("op_1410_cast_fp16")];
            tensor<int32, [2]> var_1411_split_sizes_0 = const()[name = string("op_1411_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1411_axis_0 = const()[name = string("op_1411_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1411_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1411_cast_fp16_1 = split(axis = var_1411_axis_0, split_sizes = var_1411_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1411_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_1413_cast_fp16 = mul(x = var_1411_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1413_cast_fp16")];
            int32 var_1415 = const()[name = string("op_1415"), val = int32(-2)];
            bool var_1416_interleave_0 = const()[name = string("op_1416_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1416_cast_fp16 = concat(axis = var_1415, interleave = var_1416_interleave_0, values = (var_1413_cast_fp16, var_1411_cast_fp16_0))[name = string("op_1416_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1417_cast_fp16 = mul(x = var_1416_cast_fp16, y = var_315_cast_fp16)[name = string("op_1417_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_21_cast_fp16 = add(x = var_1410_cast_fp16, y = var_1417_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1423_cast_fp16 = mul(x = var_1399_cast_fp16, y = var_308_cast_fp16)[name = string("op_1423_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1424_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = var_1399_cast_fp16)[name = string("op_1424_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_1426_cast_fp16 = mul(x = var_1424_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1426_cast_fp16")];
            int32 var_1428 = const()[name = string("op_1428"), val = int32(-2)];
            bool var_1429_interleave_0 = const()[name = string("op_1429_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1429_cast_fp16 = concat(axis = var_1428, interleave = var_1429_interleave_0, values = (var_1426_cast_fp16, var_1424_cast_fp16_0))[name = string("op_1429_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1430_cast_fp16 = mul(x = var_1429_cast_fp16, y = var_315_cast_fp16)[name = string("op_1430_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_35_cast_fp16 = add(x = var_1423_cast_fp16, y = var_1430_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (expand_dims_36, expand_dims_37, position_id, expand_dims_39))[name = string("concat_41")];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> concat_42_values1_0 = const()[name = string("concat_42_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_42_values3_0 = const()[name = string("concat_42_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_42_axis_0 = const()[name = string("concat_42_axis_0"), val = int32(0)];
            bool concat_42_interleave_0 = const()[name = string("concat_42_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_42 = concat(axis = concat_42_axis_0, interleave = concat_42_interleave_0, values = (expand_dims_40, concat_42_values1_0, cache_position_end, concat_42_values3_0))[name = string("concat_42")];
            tensor<int32, [4]> key_states_37_perm_0 = const()[name = string("key_states_37_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_37_cast_fp16 = transpose(perm = key_states_37_perm_0, x = key_states_35_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = key_states_37_cast_fp16, x = coreml_update_state_4)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_6_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_6 = read_state(input = key_cache)[name = string("coreml_update_state_6")];
            tensor<int32, [4]> value_states_21_perm_0 = const()[name = string("value_states_21_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_21_cast_fp16 = transpose(perm = value_states_21_perm_0, x = var_1406_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_41, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_42, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = value_states_21_cast_fp16, x = coreml_update_state_5)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_7_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_7 = read_state(input = value_cache)[name = string("coreml_update_state_7")];
            tensor<int32, [4]> var_1500_begin_0 = const()[name = string("op_1500_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1500_end_0 = const()[name = string("op_1500_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1500_end_mask_0 = const()[name = string("op_1500_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1500_cast_fp16 = slice_by_index(begin = var_1500_begin_0, end = var_1500_end_0, end_mask = var_1500_end_mask_0, x = coreml_update_state_6)[name = string("op_1500_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1503_axis_0 = const()[name = string("op_1503_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1503_cast_fp16_1 = split(axis = var_1503_axis_0, split_sizes = tile_6, x = var_1500_cast_fp16)[name = string("op_1503_cast_fp16")];
            tensor<int32, [4]> var_1510_begin_0 = const()[name = string("op_1510_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1510_end_0 = const()[name = string("op_1510_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1510_end_mask_0 = const()[name = string("op_1510_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1510_cast_fp16 = slice_by_index(begin = var_1510_begin_0, end = var_1510_end_0, end_mask = var_1510_end_mask_0, x = coreml_update_state_7)[name = string("op_1510_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1513_axis_0 = const()[name = string("op_1513_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1513_cast_fp16_1 = split(axis = var_1513_axis_0, split_sizes = tile_7, x = var_1510_cast_fp16)[name = string("op_1513_cast_fp16")];
            tensor<int32, [2]> var_1516_split_sizes_0 = const()[name = string("op_1516_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1516_axis_0 = const()[name = string("op_1516_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1516_0, tensor<fp16, [1, 8, 128, 1]> var_1516_1 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1516")];
            bool attn_weights_49_transpose_x_0 = const()[name = string("attn_weights_49_transpose_x_0"), val = bool(false)];
            bool attn_weights_49_transpose_y_0 = const()[name = string("attn_weights_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_0, transpose_y = attn_weights_49_transpose_y_0, x = var_1503_cast_fp16_0, y = var_1516_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_1519_to_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_53_cast_fp16 = add(x = attn_weights_51_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_53_cast_fp16")];
            int32 var_1523 = const()[name = string("op_1523"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_55_cast_fp16 = softmax(axis = var_1523, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1529_transpose_x_1 = const()[name = string("op_1529_transpose_x_1"), val = bool(true)];
            bool var_1529_transpose_y_1 = const()[name = string("op_1529_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1529_cast_fp16 = matmul(transpose_x = var_1529_transpose_x_1, transpose_y = var_1529_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1513_cast_fp16_0)[name = string("op_1529_cast_fp16")];
            bool attn_weights_57_transpose_x_0 = const()[name = string("attn_weights_57_transpose_x_0"), val = bool(false)];
            bool attn_weights_57_transpose_y_0 = const()[name = string("attn_weights_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_57_cast_fp16 = matmul(transpose_x = attn_weights_57_transpose_x_0, transpose_y = attn_weights_57_transpose_y_0, x = var_1503_cast_fp16_1, y = var_1516_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1531_to_fp16 = const()[name = string("op_1531_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_59_cast_fp16 = mul(x = attn_weights_57_cast_fp16, y = var_1531_to_fp16)[name = string("attn_weights_59_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_61_cast_fp16 = add(x = attn_weights_59_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_61_cast_fp16")];
            int32 var_1535 = const()[name = string("op_1535"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_63_cast_fp16 = softmax(axis = var_1535, x = attn_weights_61_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            bool attn_output_25_transpose_x_1 = const()[name = string("attn_output_25_transpose_x_1"), val = bool(true)];
            bool attn_output_25_transpose_y_1 = const()[name = string("attn_output_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_1, transpose_y = attn_output_25_transpose_y_1, x = attn_weights_63_cast_fp16, y = var_1513_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1543 = const()[name = string("op_1543"), val = int32(1)];
            bool attn_output_27_interleave_0 = const()[name = string("attn_output_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_27_cast_fp16 = concat(axis = var_1543, interleave = attn_output_27_interleave_0, values = (var_1529_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1547_perm_0 = const()[name = string("op_1547_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_1547_cast_fp16 = transpose(perm = var_1547_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1547_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369560512)))];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_33_cast_fp16 = conv(dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = attn_output_31_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_35_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1580_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1580_cast_fp16")];
            int32 var_1578 = const()[name = string("op_1578"), val = int32(1)];
            bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_29_cast_fp16 = concat(axis = var_1578, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1580_cast_fp16))[name = string("doubled_29_cast_fp16")];
            tensor<int32, [1]> out_15_axes_0 = const()[name = string("out_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_15_gamma_0_to_fp16 = const()[name = string("out_15_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377949184)))];
            fp16 var_1590_to_fp16 = const()[name = string("op_1590_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_1590_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1601_split_sizes_0 = const()[name = string("op_1601_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1601_axis_0 = const()[name = string("op_1601_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1601_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1601_cast_fp16_1 = split(axis = var_1601_axis_0, split_sizes = var_1601_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1601_cast_fp16")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_3_mlp_gate_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1618_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1618_cast_fp16")];
            tensor<int32, [2]> var_1624_strides_0 = const()[name = string("op_1624_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1624_pad_type_0 = const()[name = string("op_1624_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1624_pad_0 = const()[name = string("op_1624_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1624_dilations_0 = const()[name = string("op_1624_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1624_groups_0 = const()[name = string("op_1624_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1624_cast_fp16 = conv(dilations = var_1624_dilations_0, groups = var_1624_groups_0, pad = var_1624_pad_0, pad_type = var_1624_pad_type_0, strides = var_1624_strides_0, weight = layers_3_mlp_up_proj_weight_cast_fp16, x = var_1601_cast_fp16_0)[name = string("op_1624_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_39_cast_fp16 = mul(x = var_1618_cast_fp16, y = var_1624_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_37_cast_fp16 = conv(dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_3_mlp_down_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_39_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1642_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1642_cast_fp16")];
            int32 var_1640 = const()[name = string("op_1640"), val = int32(1)];
            bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_33_cast_fp16 = concat(axis = var_1640, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1642_cast_fp16))[name = string("doubled_33_cast_fp16")];
            tensor<int32, [1]> out_17_axes_0 = const()[name = string("out_17_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_17_gamma_0_to_fp16 = const()[name = string("out_17_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377957440)))];
            fp16 var_1652_to_fp16 = const()[name = string("op_1652_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1652_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1663_split_sizes_0 = const()[name = string("op_1663_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1663_axis_0 = const()[name = string("op_1663_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1663_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1663_cast_fp16_1 = split(axis = var_1663_axis_0, split_sizes = var_1663_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1663_cast_fp16")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = layers_4_self_attn_q_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = layers_4_self_attn_k_proj_weight_cast_fp16, x = var_1663_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377965696)))];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = var_1663_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_41_cast_fp16 = reshape(shape = concat_48x, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1720_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1720_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_1727_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1731_cast_fp16 = mul(x = x_41_cast_fp16, y = var_308_cast_fp16)[name = string("op_1731_cast_fp16")];
            tensor<int32, [2]> var_1732_split_sizes_0 = const()[name = string("op_1732_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1732_axis_0 = const()[name = string("op_1732_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1732_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1732_cast_fp16_1 = split(axis = var_1732_axis_0, split_sizes = var_1732_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1732_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_1734_cast_fp16 = mul(x = var_1732_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1734_cast_fp16")];
            int32 var_1736 = const()[name = string("op_1736"), val = int32(-2)];
            bool var_1737_interleave_0 = const()[name = string("op_1737_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1737_cast_fp16 = concat(axis = var_1736, interleave = var_1737_interleave_0, values = (var_1734_cast_fp16, var_1732_cast_fp16_0))[name = string("op_1737_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1738_cast_fp16 = mul(x = var_1737_cast_fp16, y = var_315_cast_fp16)[name = string("op_1738_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_27_cast_fp16 = add(x = var_1731_cast_fp16, y = var_1738_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1744_cast_fp16 = mul(x = var_1720_cast_fp16, y = var_308_cast_fp16)[name = string("op_1744_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1745_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = var_1720_cast_fp16)[name = string("op_1745_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_1747_cast_fp16 = mul(x = var_1745_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1747_cast_fp16")];
            int32 var_1749 = const()[name = string("op_1749"), val = int32(-2)];
            bool var_1750_interleave_0 = const()[name = string("op_1750_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1750_cast_fp16 = concat(axis = var_1749, interleave = var_1750_interleave_0, values = (var_1747_cast_fp16, var_1745_cast_fp16_0))[name = string("op_1750_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1751_cast_fp16 = mul(x = var_1750_cast_fp16, y = var_315_cast_fp16)[name = string("op_1751_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_45_cast_fp16 = add(x = var_1744_cast_fp16, y = var_1751_cast_fp16)[name = string("key_states_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (expand_dims_48, expand_dims_49, position_id, expand_dims_51))[name = string("concat_53")];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> concat_54_values1_0 = const()[name = string("concat_54_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_54_values3_0 = const()[name = string("concat_54_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)];
            bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (expand_dims_52, concat_54_values1_0, cache_position_end, concat_54_values3_0))[name = string("concat_54")];
            tensor<int32, [4]> key_states_47_perm_0 = const()[name = string("key_states_47_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_47_cast_fp16 = transpose(perm = key_states_47_perm_0, x = key_states_45_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = key_states_47_cast_fp16, x = coreml_update_state_6)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_8_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_8 = read_state(input = key_cache)[name = string("coreml_update_state_8")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = string("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_1727_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_53, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_54, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = value_states_27_cast_fp16, x = coreml_update_state_7)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_9_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_9 = read_state(input = value_cache)[name = string("coreml_update_state_9")];
            tensor<int32, [4]> var_1821_begin_0 = const()[name = string("op_1821_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1821_end_0 = const()[name = string("op_1821_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1821_end_mask_0 = const()[name = string("op_1821_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1821_cast_fp16 = slice_by_index(begin = var_1821_begin_0, end = var_1821_end_0, end_mask = var_1821_end_mask_0, x = coreml_update_state_8)[name = string("op_1821_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1824_axis_0 = const()[name = string("op_1824_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1824_cast_fp16_1 = split(axis = var_1824_axis_0, split_sizes = tile_8, x = var_1821_cast_fp16)[name = string("op_1824_cast_fp16")];
            tensor<int32, [4]> var_1831_begin_0 = const()[name = string("op_1831_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1831_end_0 = const()[name = string("op_1831_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1831_end_mask_0 = const()[name = string("op_1831_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1831_cast_fp16 = slice_by_index(begin = var_1831_begin_0, end = var_1831_end_0, end_mask = var_1831_end_mask_0, x = coreml_update_state_9)[name = string("op_1831_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1834_axis_0 = const()[name = string("op_1834_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1834_cast_fp16_1 = split(axis = var_1834_axis_0, split_sizes = tile_9, x = var_1831_cast_fp16)[name = string("op_1834_cast_fp16")];
            tensor<int32, [2]> var_1837_split_sizes_0 = const()[name = string("op_1837_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1837_axis_0 = const()[name = string("op_1837_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1837_0, tensor<fp16, [1, 8, 128, 1]> var_1837_1 = split(axis = var_1837_axis_0, split_sizes = var_1837_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1837")];
            bool attn_weights_65_transpose_x_0 = const()[name = string("attn_weights_65_transpose_x_0"), val = bool(false)];
            bool attn_weights_65_transpose_y_0 = const()[name = string("attn_weights_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_65_cast_fp16 = matmul(transpose_x = attn_weights_65_transpose_x_0, transpose_y = attn_weights_65_transpose_y_0, x = var_1824_cast_fp16_0, y = var_1837_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1840_to_fp16 = const()[name = string("op_1840_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_67_cast_fp16 = mul(x = attn_weights_65_cast_fp16, y = var_1840_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            int32 var_1844 = const()[name = string("op_1844"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_71_cast_fp16 = softmax(axis = var_1844, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1850_transpose_x_1 = const()[name = string("op_1850_transpose_x_1"), val = bool(true)];
            bool var_1850_transpose_y_1 = const()[name = string("op_1850_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1850_cast_fp16 = matmul(transpose_x = var_1850_transpose_x_1, transpose_y = var_1850_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1834_cast_fp16_0)[name = string("op_1850_cast_fp16")];
            bool attn_weights_73_transpose_x_0 = const()[name = string("attn_weights_73_transpose_x_0"), val = bool(false)];
            bool attn_weights_73_transpose_y_0 = const()[name = string("attn_weights_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_0, transpose_y = attn_weights_73_transpose_y_0, x = var_1824_cast_fp16_1, y = var_1837_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1852_to_fp16 = const()[name = string("op_1852_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1852_to_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_77_cast_fp16 = add(x = attn_weights_75_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_77_cast_fp16")];
            int32 var_1856 = const()[name = string("op_1856"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_79_cast_fp16 = softmax(axis = var_1856, x = attn_weights_77_cast_fp16)[name = string("attn_weights_79_cast_fp16")];
            bool attn_output_33_transpose_x_1 = const()[name = string("attn_output_33_transpose_x_1"), val = bool(true)];
            bool attn_output_33_transpose_y_1 = const()[name = string("attn_output_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_1, transpose_y = attn_output_33_transpose_y_1, x = attn_weights_79_cast_fp16, y = var_1834_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_1864 = const()[name = string("op_1864"), val = int32(1)];
            bool attn_output_35_interleave_0 = const()[name = string("attn_output_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_35_cast_fp16 = concat(axis = var_1864, interleave = attn_output_35_interleave_0, values = (var_1850_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_1868_perm_0 = const()[name = string("op_1868_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_1868_cast_fp16 = transpose(perm = var_1868_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_1868_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_43_cast_fp16 = conv(dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_4_self_attn_o_proj_weight_cast_fp16, x = attn_output_39_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_45_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1901_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_1901_cast_fp16")];
            int32 var_1899 = const()[name = string("op_1899"), val = int32(1)];
            bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_37_cast_fp16 = concat(axis = var_1899, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_1901_cast_fp16))[name = string("doubled_37_cast_fp16")];
            tensor<int32, [1]> out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_19_gamma_0_to_fp16 = const()[name = string("out_19_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379014336)))];
            fp16 var_1911_to_fp16 = const()[name = string("op_1911_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1911_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_1922_split_sizes_0 = const()[name = string("op_1922_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1922_axis_0 = const()[name = string("op_1922_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1922_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1922_cast_fp16_1 = split(axis = var_1922_axis_0, split_sizes = var_1922_split_sizes_0, x = out_19_cast_fp16)[name = string("op_1922_cast_fp16")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_4_mlp_gate_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1939_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<int32, [2]> var_1945_strides_0 = const()[name = string("op_1945_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1945_pad_type_0 = const()[name = string("op_1945_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1945_pad_0 = const()[name = string("op_1945_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1945_dilations_0 = const()[name = string("op_1945_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1945_groups_0 = const()[name = string("op_1945_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1945_cast_fp16 = conv(dilations = var_1945_dilations_0, groups = var_1945_groups_0, pad = var_1945_pad_0, pad_type = var_1945_pad_type_0, strides = var_1945_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_1922_cast_fp16_0)[name = string("op_1945_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_49_cast_fp16 = mul(x = var_1939_cast_fp16, y = var_1945_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_47_cast_fp16 = conv(dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_4_mlp_down_proj_weight_cast_fp16, x = x_49_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_49_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_1963_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_1963_cast_fp16")];
            int32 var_1961 = const()[name = string("op_1961"), val = int32(1)];
            bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_41_cast_fp16 = concat(axis = var_1961, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_1963_cast_fp16))[name = string("doubled_41_cast_fp16")];
            tensor<int32, [1]> out_21_axes_0 = const()[name = string("out_21_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_21_gamma_0_to_fp16 = const()[name = string("out_21_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379022592)))];
            fp16 var_1973_to_fp16 = const()[name = string("op_1973_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1973_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_1984_split_sizes_0 = const()[name = string("op_1984_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1984_axis_0 = const()[name = string("op_1984_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1984_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1984_cast_fp16_1 = split(axis = var_1984_axis_0, split_sizes = var_1984_split_sizes_0, x = out_21_cast_fp16)[name = string("op_1984_cast_fp16")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_31_pad_0 = const()[name = string("query_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_31_dilations_0 = const()[name = string("query_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_31_groups_0 = const()[name = string("query_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_31_cast_fp16 = conv(dilations = query_states_31_dilations_0, groups = query_states_31_groups_0, pad = query_states_31_pad_0, pad_type = query_states_31_pad_type_0, strides = query_states_31_strides_0, weight = layers_5_self_attn_q_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_51_pad_0 = const()[name = string("key_states_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_51_dilations_0 = const()[name = string("key_states_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_51_groups_0 = const()[name = string("key_states_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_51_cast_fp16 = conv(dilations = key_states_51_dilations_0, groups = key_states_51_groups_0, pad = key_states_51_pad_0, pad_type = key_states_51_pad_type_0, strides = key_states_51_strides_0, weight = layers_5_self_attn_k_proj_weight_cast_fp16, x = var_1984_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379030848)))];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_31_pad_0 = const()[name = string("value_states_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_31_dilations_0 = const()[name = string("value_states_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_31_groups_0 = const()[name = string("value_states_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_31_cast_fp16 = conv(dilations = value_states_31_dilations_0, groups = value_states_31_groups_0, pad = value_states_31_pad_0, pad_type = value_states_31_pad_type_0, strides = value_states_31_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = var_1984_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_51_cast_fp16 = reshape(shape = concat_60x, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2041_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2041_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2048_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2048_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2052_cast_fp16 = mul(x = x_51_cast_fp16, y = var_308_cast_fp16)[name = string("op_2052_cast_fp16")];
            tensor<int32, [2]> var_2053_split_sizes_0 = const()[name = string("op_2053_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2053_axis_0 = const()[name = string("op_2053_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2053_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2053_cast_fp16_1 = split(axis = var_2053_axis_0, split_sizes = var_2053_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2053_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_2055_cast_fp16 = mul(x = var_2053_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2055_cast_fp16")];
            int32 var_2057 = const()[name = string("op_2057"), val = int32(-2)];
            bool var_2058_interleave_0 = const()[name = string("op_2058_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2058_cast_fp16 = concat(axis = var_2057, interleave = var_2058_interleave_0, values = (var_2055_cast_fp16, var_2053_cast_fp16_0))[name = string("op_2058_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2059_cast_fp16 = mul(x = var_2058_cast_fp16, y = var_315_cast_fp16)[name = string("op_2059_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_33_cast_fp16 = add(x = var_2052_cast_fp16, y = var_2059_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2065_cast_fp16 = mul(x = var_2041_cast_fp16, y = var_308_cast_fp16)[name = string("op_2065_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2066_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = var_2041_cast_fp16)[name = string("op_2066_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_2068_cast_fp16 = mul(x = var_2066_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2068_cast_fp16")];
            int32 var_2070 = const()[name = string("op_2070"), val = int32(-2)];
            bool var_2071_interleave_0 = const()[name = string("op_2071_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2071_cast_fp16 = concat(axis = var_2070, interleave = var_2071_interleave_0, values = (var_2068_cast_fp16, var_2066_cast_fp16_0))[name = string("op_2071_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2072_cast_fp16 = mul(x = var_2071_cast_fp16, y = var_315_cast_fp16)[name = string("op_2072_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_55_cast_fp16 = add(x = var_2065_cast_fp16, y = var_2072_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (expand_dims_60, expand_dims_61, position_id, expand_dims_63))[name = string("concat_65")];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> concat_66_values1_0 = const()[name = string("concat_66_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_66_values3_0 = const()[name = string("concat_66_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_66_axis_0 = const()[name = string("concat_66_axis_0"), val = int32(0)];
            bool concat_66_interleave_0 = const()[name = string("concat_66_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_66 = concat(axis = concat_66_axis_0, interleave = concat_66_interleave_0, values = (expand_dims_64, concat_66_values1_0, cache_position_end, concat_66_values3_0))[name = string("concat_66")];
            tensor<int32, [4]> key_states_57_perm_0 = const()[name = string("key_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_57_cast_fp16 = transpose(perm = key_states_57_perm_0, x = key_states_55_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = key_states_57_cast_fp16, x = coreml_update_state_8)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_10_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_10 = read_state(input = key_cache)[name = string("coreml_update_state_10")];
            tensor<int32, [4]> value_states_33_perm_0 = const()[name = string("value_states_33_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_33_cast_fp16 = transpose(perm = value_states_33_perm_0, x = var_2048_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_65, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_66, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = value_states_33_cast_fp16, x = coreml_update_state_9)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_11_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_11 = read_state(input = value_cache)[name = string("coreml_update_state_11")];
            tensor<int32, [4]> var_2142_begin_0 = const()[name = string("op_2142_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2142_end_0 = const()[name = string("op_2142_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2142_end_mask_0 = const()[name = string("op_2142_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2142_cast_fp16 = slice_by_index(begin = var_2142_begin_0, end = var_2142_end_0, end_mask = var_2142_end_mask_0, x = coreml_update_state_10)[name = string("op_2142_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2145_axis_0 = const()[name = string("op_2145_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2145_cast_fp16_1 = split(axis = var_2145_axis_0, split_sizes = tile_10, x = var_2142_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [4]> var_2152_begin_0 = const()[name = string("op_2152_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2152_end_0 = const()[name = string("op_2152_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2152_end_mask_0 = const()[name = string("op_2152_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2152_cast_fp16 = slice_by_index(begin = var_2152_begin_0, end = var_2152_end_0, end_mask = var_2152_end_mask_0, x = coreml_update_state_11)[name = string("op_2152_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2155_axis_0 = const()[name = string("op_2155_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2155_cast_fp16_1 = split(axis = var_2155_axis_0, split_sizes = tile_11, x = var_2152_cast_fp16)[name = string("op_2155_cast_fp16")];
            tensor<int32, [2]> var_2158_split_sizes_0 = const()[name = string("op_2158_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2158_axis_0 = const()[name = string("op_2158_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2158_0, tensor<fp16, [1, 8, 128, 1]> var_2158_1 = split(axis = var_2158_axis_0, split_sizes = var_2158_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2158")];
            bool attn_weights_81_transpose_x_0 = const()[name = string("attn_weights_81_transpose_x_0"), val = bool(false)];
            bool attn_weights_81_transpose_y_0 = const()[name = string("attn_weights_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_81_cast_fp16 = matmul(transpose_x = attn_weights_81_transpose_x_0, transpose_y = attn_weights_81_transpose_y_0, x = var_2145_cast_fp16_0, y = var_2158_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2161_to_fp16 = const()[name = string("op_2161_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_83_cast_fp16 = mul(x = attn_weights_81_cast_fp16, y = var_2161_to_fp16)[name = string("attn_weights_83_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_85_cast_fp16 = add(x = attn_weights_83_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_85_cast_fp16")];
            int32 var_2165 = const()[name = string("op_2165"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_87_cast_fp16 = softmax(axis = var_2165, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2171_transpose_x_1 = const()[name = string("op_2171_transpose_x_1"), val = bool(true)];
            bool var_2171_transpose_y_1 = const()[name = string("op_2171_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2171_cast_fp16 = matmul(transpose_x = var_2171_transpose_x_1, transpose_y = var_2171_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2155_cast_fp16_0)[name = string("op_2171_cast_fp16")];
            bool attn_weights_89_transpose_x_0 = const()[name = string("attn_weights_89_transpose_x_0"), val = bool(false)];
            bool attn_weights_89_transpose_y_0 = const()[name = string("attn_weights_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_89_cast_fp16 = matmul(transpose_x = attn_weights_89_transpose_x_0, transpose_y = attn_weights_89_transpose_y_0, x = var_2145_cast_fp16_1, y = var_2158_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2173_to_fp16 = const()[name = string("op_2173_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_91_cast_fp16 = mul(x = attn_weights_89_cast_fp16, y = var_2173_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            int32 var_2177 = const()[name = string("op_2177"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_95_cast_fp16 = softmax(axis = var_2177, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")];
            bool attn_output_41_transpose_x_1 = const()[name = string("attn_output_41_transpose_x_1"), val = bool(true)];
            bool attn_output_41_transpose_y_1 = const()[name = string("attn_output_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_1, transpose_y = attn_output_41_transpose_y_1, x = attn_weights_95_cast_fp16, y = var_2155_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2185 = const()[name = string("op_2185"), val = int32(1)];
            bool attn_output_43_interleave_0 = const()[name = string("attn_output_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_43_cast_fp16 = concat(axis = var_2185, interleave = attn_output_43_interleave_0, values = (var_2171_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2189_perm_0 = const()[name = string("op_2189_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_2189_cast_fp16 = transpose(perm = var_2189_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2189_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_53_cast_fp16 = conv(dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_5_self_attn_o_proj_weight_cast_fp16, x = attn_output_47_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_55_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2222_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2222_cast_fp16")];
            int32 var_2220 = const()[name = string("op_2220"), val = int32(1)];
            bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_45_cast_fp16 = concat(axis = var_2220, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2222_cast_fp16))[name = string("doubled_45_cast_fp16")];
            tensor<int32, [1]> out_23_axes_0 = const()[name = string("out_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_23_gamma_0_to_fp16 = const()[name = string("out_23_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380079488)))];
            fp16 var_2232_to_fp16 = const()[name = string("op_2232_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_2232_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2243_split_sizes_0 = const()[name = string("op_2243_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2243_axis_0 = const()[name = string("op_2243_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2243_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2243_cast_fp16_1 = split(axis = var_2243_axis_0, split_sizes = var_2243_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2243_cast_fp16")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_5_mlp_gate_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2260_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<int32, [2]> var_2266_strides_0 = const()[name = string("op_2266_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2266_pad_type_0 = const()[name = string("op_2266_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2266_pad_0 = const()[name = string("op_2266_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2266_dilations_0 = const()[name = string("op_2266_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2266_groups_0 = const()[name = string("op_2266_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2266_cast_fp16 = conv(dilations = var_2266_dilations_0, groups = var_2266_groups_0, pad = var_2266_pad_0, pad_type = var_2266_pad_type_0, strides = var_2266_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2243_cast_fp16_0)[name = string("op_2266_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_59_cast_fp16 = mul(x = var_2260_cast_fp16, y = var_2266_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_57_cast_fp16 = conv(dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_5_mlp_down_proj_weight_cast_fp16, x = x_59_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_59_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2284_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2284_cast_fp16")];
            int32 var_2282 = const()[name = string("op_2282"), val = int32(1)];
            bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_49_cast_fp16 = concat(axis = var_2282, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2284_cast_fp16))[name = string("doubled_49_cast_fp16")];
            tensor<int32, [1]> out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_25_gamma_0_to_fp16 = const()[name = string("out_25_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380087744)))];
            fp16 var_2294_to_fp16 = const()[name = string("op_2294_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_2294_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2305_split_sizes_0 = const()[name = string("op_2305_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2305_axis_0 = const()[name = string("op_2305_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2305_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2305_cast_fp16_1 = split(axis = var_2305_axis_0, split_sizes = var_2305_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = layers_6_self_attn_q_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_61_pad_0 = const()[name = string("key_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_61_dilations_0 = const()[name = string("key_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_61_groups_0 = const()[name = string("key_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_61_cast_fp16 = conv(dilations = key_states_61_dilations_0, groups = key_states_61_groups_0, pad = key_states_61_pad_0, pad_type = key_states_61_pad_type_0, strides = key_states_61_strides_0, weight = layers_6_self_attn_k_proj_weight_cast_fp16, x = var_2305_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380096000)))];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = var_2305_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_61_cast_fp16 = reshape(shape = concat_72x, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2362_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2362_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2369_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2369_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2373_cast_fp16 = mul(x = x_61_cast_fp16, y = var_308_cast_fp16)[name = string("op_2373_cast_fp16")];
            tensor<int32, [2]> var_2374_split_sizes_0 = const()[name = string("op_2374_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2374_axis_0 = const()[name = string("op_2374_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2374_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2374_cast_fp16_1 = split(axis = var_2374_axis_0, split_sizes = var_2374_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2376_cast_fp16")];
            int32 var_2378 = const()[name = string("op_2378"), val = int32(-2)];
            bool var_2379_interleave_0 = const()[name = string("op_2379_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2379_cast_fp16 = concat(axis = var_2378, interleave = var_2379_interleave_0, values = (var_2376_cast_fp16, var_2374_cast_fp16_0))[name = string("op_2379_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2380_cast_fp16 = mul(x = var_2379_cast_fp16, y = var_315_cast_fp16)[name = string("op_2380_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_39_cast_fp16 = add(x = var_2373_cast_fp16, y = var_2380_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2386_cast_fp16 = mul(x = var_2362_cast_fp16, y = var_308_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2387_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = var_2362_cast_fp16)[name = string("op_2387_cast_fp16")];
            fp16 const_65_promoted_to_fp16 = const()[name = string("const_65_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_2389_cast_fp16 = mul(x = var_2387_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2389_cast_fp16")];
            int32 var_2391 = const()[name = string("op_2391"), val = int32(-2)];
            bool var_2392_interleave_0 = const()[name = string("op_2392_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2392_cast_fp16 = concat(axis = var_2391, interleave = var_2392_interleave_0, values = (var_2389_cast_fp16, var_2387_cast_fp16_0))[name = string("op_2392_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2393_cast_fp16 = mul(x = var_2392_cast_fp16, y = var_315_cast_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_65_cast_fp16 = add(x = var_2386_cast_fp16, y = var_2393_cast_fp16)[name = string("key_states_65_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (expand_dims_72, expand_dims_73, position_id, expand_dims_75))[name = string("concat_77")];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> concat_78_values1_0 = const()[name = string("concat_78_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_78_values3_0 = const()[name = string("concat_78_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_78_axis_0 = const()[name = string("concat_78_axis_0"), val = int32(0)];
            bool concat_78_interleave_0 = const()[name = string("concat_78_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_78 = concat(axis = concat_78_axis_0, interleave = concat_78_interleave_0, values = (expand_dims_76, concat_78_values1_0, cache_position_end, concat_78_values3_0))[name = string("concat_78")];
            tensor<int32, [4]> key_states_67_perm_0 = const()[name = string("key_states_67_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_67_cast_fp16 = transpose(perm = key_states_67_perm_0, x = key_states_65_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = key_states_67_cast_fp16, x = coreml_update_state_10)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_12_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_12 = read_state(input = key_cache)[name = string("coreml_update_state_12")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = string("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_2369_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_77, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_78, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = value_states_39_cast_fp16, x = coreml_update_state_11)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_13_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_13 = read_state(input = value_cache)[name = string("coreml_update_state_13")];
            tensor<int32, [4]> var_2463_begin_0 = const()[name = string("op_2463_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2463_end_0 = const()[name = string("op_2463_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2463_end_mask_0 = const()[name = string("op_2463_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2463_cast_fp16 = slice_by_index(begin = var_2463_begin_0, end = var_2463_end_0, end_mask = var_2463_end_mask_0, x = coreml_update_state_12)[name = string("op_2463_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2466_axis_0 = const()[name = string("op_2466_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2466_cast_fp16_1 = split(axis = var_2466_axis_0, split_sizes = tile_12, x = var_2463_cast_fp16)[name = string("op_2466_cast_fp16")];
            tensor<int32, [4]> var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, x = coreml_update_state_13)[name = string("op_2473_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2476_axis_0 = const()[name = string("op_2476_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2476_cast_fp16_1 = split(axis = var_2476_axis_0, split_sizes = tile_13, x = var_2473_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<int32, [2]> var_2479_split_sizes_0 = const()[name = string("op_2479_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2479_axis_0 = const()[name = string("op_2479_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2479_0, tensor<fp16, [1, 8, 128, 1]> var_2479_1 = split(axis = var_2479_axis_0, split_sizes = var_2479_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2479")];
            bool attn_weights_97_transpose_x_0 = const()[name = string("attn_weights_97_transpose_x_0"), val = bool(false)];
            bool attn_weights_97_transpose_y_0 = const()[name = string("attn_weights_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_0, transpose_y = attn_weights_97_transpose_y_0, x = var_2466_cast_fp16_0, y = var_2479_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2482_to_fp16 = const()[name = string("op_2482_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_2482_to_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_101_cast_fp16 = add(x = attn_weights_99_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_101_cast_fp16")];
            int32 var_2486 = const()[name = string("op_2486"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_103_cast_fp16 = softmax(axis = var_2486, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2492_transpose_x_1 = const()[name = string("op_2492_transpose_x_1"), val = bool(true)];
            bool var_2492_transpose_y_1 = const()[name = string("op_2492_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2492_cast_fp16 = matmul(transpose_x = var_2492_transpose_x_1, transpose_y = var_2492_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2476_cast_fp16_0)[name = string("op_2492_cast_fp16")];
            bool attn_weights_105_transpose_x_0 = const()[name = string("attn_weights_105_transpose_x_0"), val = bool(false)];
            bool attn_weights_105_transpose_y_0 = const()[name = string("attn_weights_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_105_cast_fp16 = matmul(transpose_x = attn_weights_105_transpose_x_0, transpose_y = attn_weights_105_transpose_y_0, x = var_2466_cast_fp16_1, y = var_2479_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2494_to_fp16 = const()[name = string("op_2494_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_107_cast_fp16 = mul(x = attn_weights_105_cast_fp16, y = var_2494_to_fp16)[name = string("attn_weights_107_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_109_cast_fp16 = add(x = attn_weights_107_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_109_cast_fp16")];
            int32 var_2498 = const()[name = string("op_2498"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_111_cast_fp16 = softmax(axis = var_2498, x = attn_weights_109_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            bool attn_output_49_transpose_x_1 = const()[name = string("attn_output_49_transpose_x_1"), val = bool(true)];
            bool attn_output_49_transpose_y_1 = const()[name = string("attn_output_49_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_1, transpose_y = attn_output_49_transpose_y_1, x = attn_weights_111_cast_fp16, y = var_2476_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2506 = const()[name = string("op_2506"), val = int32(1)];
            bool attn_output_51_interleave_0 = const()[name = string("attn_output_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_51_cast_fp16 = concat(axis = var_2506, interleave = attn_output_51_interleave_0, values = (var_2492_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2510_perm_0 = const()[name = string("op_2510_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_2510_cast_fp16 = transpose(perm = var_2510_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2510_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_63_cast_fp16 = conv(dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_6_self_attn_o_proj_weight_cast_fp16, x = attn_output_55_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_65_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2543_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2543_cast_fp16")];
            int32 var_2541 = const()[name = string("op_2541"), val = int32(1)];
            bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_53_cast_fp16 = concat(axis = var_2541, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2543_cast_fp16))[name = string("doubled_53_cast_fp16")];
            tensor<int32, [1]> out_27_axes_0 = const()[name = string("out_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_27_gamma_0_to_fp16 = const()[name = string("out_27_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381144640)))];
            fp16 var_2553_to_fp16 = const()[name = string("op_2553_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_2553_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2564_split_sizes_0 = const()[name = string("op_2564_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2564_axis_0 = const()[name = string("op_2564_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2564_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2564_cast_fp16_1 = split(axis = var_2564_axis_0, split_sizes = var_2564_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_6_mlp_gate_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2581_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [2]> var_2587_strides_0 = const()[name = string("op_2587_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2587_pad_type_0 = const()[name = string("op_2587_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2587_pad_0 = const()[name = string("op_2587_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2587_dilations_0 = const()[name = string("op_2587_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2587_groups_0 = const()[name = string("op_2587_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2587_cast_fp16 = conv(dilations = var_2587_dilations_0, groups = var_2587_groups_0, pad = var_2587_pad_0, pad_type = var_2587_pad_type_0, strides = var_2587_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2564_cast_fp16_0)[name = string("op_2587_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_69_cast_fp16 = mul(x = var_2581_cast_fp16, y = var_2587_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381152896)))];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_67_pad_0 = const()[name = string("hidden_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_67_dilations_0 = const()[name = string("hidden_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_67_groups_0 = const()[name = string("hidden_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_67_cast_fp16 = conv(dilations = hidden_states_67_dilations_0, groups = hidden_states_67_groups_0, pad = hidden_states_67_pad_0, pad_type = hidden_states_67_pad_type_0, strides = hidden_states_67_strides_0, weight = layers_6_mlp_down_proj_weight_to_fp16, x = x_69_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_69_cast_fp16 = add(x = hidden_states_65_cast_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_states_69_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2605_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2605_cast_fp16")];
            int32 var_2603 = const()[name = string("op_2603"), val = int32(1)];
            bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_57_cast_fp16 = concat(axis = var_2603, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2605_cast_fp16))[name = string("doubled_57_cast_fp16")];
            tensor<int32, [1]> out_29_axes_0 = const()[name = string("out_29_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_29_gamma_0_to_fp16 = const()[name = string("out_29_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406318784)))];
            fp16 var_2615_to_fp16 = const()[name = string("op_2615_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_2615_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2626_split_sizes_0 = const()[name = string("op_2626_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2626_axis_0 = const()[name = string("op_2626_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2626_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2626_cast_fp16_1 = split(axis = var_2626_axis_0, split_sizes = var_2626_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2626_cast_fp16")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_43_pad_0 = const()[name = string("query_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_43_dilations_0 = const()[name = string("query_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_43_groups_0 = const()[name = string("query_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_43_cast_fp16 = conv(dilations = query_states_43_dilations_0, groups = query_states_43_groups_0, pad = query_states_43_pad_0, pad_type = query_states_43_pad_type_0, strides = query_states_43_strides_0, weight = layers_7_self_attn_q_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_71_pad_0 = const()[name = string("key_states_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_71_dilations_0 = const()[name = string("key_states_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_71_groups_0 = const()[name = string("key_states_71_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_71_cast_fp16 = conv(dilations = key_states_71_dilations_0, groups = key_states_71_groups_0, pad = key_states_71_pad_0, pad_type = key_states_71_pad_type_0, strides = key_states_71_strides_0, weight = layers_7_self_attn_k_proj_weight_cast_fp16, x = var_2626_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406327040)))];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_43_pad_0 = const()[name = string("value_states_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_43_dilations_0 = const()[name = string("value_states_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_43_groups_0 = const()[name = string("value_states_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_43_cast_fp16 = conv(dilations = value_states_43_dilations_0, groups = value_states_43_groups_0, pad = value_states_43_pad_0, pad_type = value_states_43_pad_type_0, strides = value_states_43_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = var_2626_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_71_cast_fp16 = reshape(shape = concat_84x, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2683_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_2690_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2690_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2694_cast_fp16 = mul(x = x_71_cast_fp16, y = var_308_cast_fp16)[name = string("op_2694_cast_fp16")];
            tensor<int32, [2]> var_2695_split_sizes_0 = const()[name = string("op_2695_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2695_axis_0 = const()[name = string("op_2695_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2695_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2695_cast_fp16_1 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2695_cast_fp16")];
            fp16 const_74_promoted_to_fp16 = const()[name = string("const_74_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_2697_cast_fp16 = mul(x = var_2695_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2697_cast_fp16")];
            int32 var_2699 = const()[name = string("op_2699"), val = int32(-2)];
            bool var_2700_interleave_0 = const()[name = string("op_2700_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2700_cast_fp16 = concat(axis = var_2699, interleave = var_2700_interleave_0, values = (var_2697_cast_fp16, var_2695_cast_fp16_0))[name = string("op_2700_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2701_cast_fp16 = mul(x = var_2700_cast_fp16, y = var_315_cast_fp16)[name = string("op_2701_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_45_cast_fp16 = add(x = var_2694_cast_fp16, y = var_2701_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2707_cast_fp16 = mul(x = var_2683_cast_fp16, y = var_308_cast_fp16)[name = string("op_2707_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2708_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = var_2683_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_2710_cast_fp16 = mul(x = var_2708_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2710_cast_fp16")];
            int32 var_2712 = const()[name = string("op_2712"), val = int32(-2)];
            bool var_2713_interleave_0 = const()[name = string("op_2713_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2713_cast_fp16 = concat(axis = var_2712, interleave = var_2713_interleave_0, values = (var_2710_cast_fp16, var_2708_cast_fp16_0))[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2714_cast_fp16 = mul(x = var_2713_cast_fp16, y = var_315_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_75_cast_fp16 = add(x = var_2707_cast_fp16, y = var_2714_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (expand_dims_84, expand_dims_85, position_id, expand_dims_87))[name = string("concat_89")];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> concat_90_values1_0 = const()[name = string("concat_90_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_90_values3_0 = const()[name = string("concat_90_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_90_axis_0 = const()[name = string("concat_90_axis_0"), val = int32(0)];
            bool concat_90_interleave_0 = const()[name = string("concat_90_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_90 = concat(axis = concat_90_axis_0, interleave = concat_90_interleave_0, values = (expand_dims_88, concat_90_values1_0, cache_position_end, concat_90_values3_0))[name = string("concat_90")];
            tensor<int32, [4]> key_states_77_perm_0 = const()[name = string("key_states_77_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_77_cast_fp16 = transpose(perm = key_states_77_perm_0, x = key_states_75_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = key_states_77_cast_fp16, x = coreml_update_state_12)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_14_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_14 = read_state(input = key_cache)[name = string("coreml_update_state_14")];
            tensor<int32, [4]> value_states_45_perm_0 = const()[name = string("value_states_45_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_45_cast_fp16 = transpose(perm = value_states_45_perm_0, x = var_2690_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_89, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_90, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = value_states_45_cast_fp16, x = coreml_update_state_13)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_15_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_15 = read_state(input = value_cache)[name = string("coreml_update_state_15")];
            tensor<int32, [4]> var_2784_begin_0 = const()[name = string("op_2784_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2784_end_0 = const()[name = string("op_2784_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2784_end_mask_0 = const()[name = string("op_2784_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2784_cast_fp16 = slice_by_index(begin = var_2784_begin_0, end = var_2784_end_0, end_mask = var_2784_end_mask_0, x = coreml_update_state_14)[name = string("op_2784_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2787_axis_0 = const()[name = string("op_2787_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2787_cast_fp16_1 = split(axis = var_2787_axis_0, split_sizes = tile_14, x = var_2784_cast_fp16)[name = string("op_2787_cast_fp16")];
            tensor<int32, [4]> var_2794_begin_0 = const()[name = string("op_2794_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2794_end_0 = const()[name = string("op_2794_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2794_end_mask_0 = const()[name = string("op_2794_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2794_cast_fp16 = slice_by_index(begin = var_2794_begin_0, end = var_2794_end_0, end_mask = var_2794_end_mask_0, x = coreml_update_state_15)[name = string("op_2794_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2797_axis_0 = const()[name = string("op_2797_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2797_cast_fp16_1 = split(axis = var_2797_axis_0, split_sizes = tile_15, x = var_2794_cast_fp16)[name = string("op_2797_cast_fp16")];
            tensor<int32, [2]> var_2800_split_sizes_0 = const()[name = string("op_2800_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2800_axis_0 = const()[name = string("op_2800_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2800_0, tensor<fp16, [1, 8, 128, 1]> var_2800_1 = split(axis = var_2800_axis_0, split_sizes = var_2800_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2800")];
            bool attn_weights_113_transpose_x_0 = const()[name = string("attn_weights_113_transpose_x_0"), val = bool(false)];
            bool attn_weights_113_transpose_y_0 = const()[name = string("attn_weights_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_113_cast_fp16 = matmul(transpose_x = attn_weights_113_transpose_x_0, transpose_y = attn_weights_113_transpose_y_0, x = var_2787_cast_fp16_0, y = var_2800_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2803_to_fp16 = const()[name = string("op_2803_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_115_cast_fp16 = mul(x = attn_weights_113_cast_fp16, y = var_2803_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            int32 var_2807 = const()[name = string("op_2807"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_119_cast_fp16 = softmax(axis = var_2807, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2813_transpose_x_1 = const()[name = string("op_2813_transpose_x_1"), val = bool(true)];
            bool var_2813_transpose_y_1 = const()[name = string("op_2813_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2813_cast_fp16 = matmul(transpose_x = var_2813_transpose_x_1, transpose_y = var_2813_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2797_cast_fp16_0)[name = string("op_2813_cast_fp16")];
            bool attn_weights_121_transpose_x_0 = const()[name = string("attn_weights_121_transpose_x_0"), val = bool(false)];
            bool attn_weights_121_transpose_y_0 = const()[name = string("attn_weights_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_0, transpose_y = attn_weights_121_transpose_y_0, x = var_2787_cast_fp16_1, y = var_2800_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2815_to_fp16 = const()[name = string("op_2815_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_2815_to_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_125_cast_fp16 = add(x = attn_weights_123_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_125_cast_fp16")];
            int32 var_2819 = const()[name = string("op_2819"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_127_cast_fp16 = softmax(axis = var_2819, x = attn_weights_125_cast_fp16)[name = string("attn_weights_127_cast_fp16")];
            bool attn_output_57_transpose_x_1 = const()[name = string("attn_output_57_transpose_x_1"), val = bool(true)];
            bool attn_output_57_transpose_y_1 = const()[name = string("attn_output_57_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_1, transpose_y = attn_output_57_transpose_y_1, x = attn_weights_127_cast_fp16, y = var_2797_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2827 = const()[name = string("op_2827"), val = int32(1)];
            bool attn_output_59_interleave_0 = const()[name = string("attn_output_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_59_cast_fp16 = concat(axis = var_2827, interleave = attn_output_59_interleave_0, values = (var_2813_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2831_perm_0 = const()[name = string("op_2831_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_2831_cast_fp16 = transpose(perm = var_2831_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2831_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_73_pad_0 = const()[name = string("hidden_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_73_dilations_0 = const()[name = string("hidden_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_73_groups_0 = const()[name = string("hidden_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_73_cast_fp16 = conv(dilations = hidden_states_73_dilations_0, groups = hidden_states_73_groups_0, pad = hidden_states_73_pad_0, pad_type = hidden_states_73_pad_type_0, strides = hidden_states_73_strides_0, weight = layers_7_self_attn_o_proj_weight_cast_fp16, x = attn_output_63_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_75_cast_fp16 = add(x = hidden_states_69_cast_fp16, y = hidden_states_73_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2864_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_2864_cast_fp16")];
            int32 var_2862 = const()[name = string("op_2862"), val = int32(1)];
            bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_61_cast_fp16 = concat(axis = var_2862, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_2864_cast_fp16))[name = string("doubled_61_cast_fp16")];
            tensor<int32, [1]> out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_31_gamma_0_to_fp16 = const()[name = string("out_31_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407375680)))];
            fp16 var_2874_to_fp16 = const()[name = string("op_2874_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_2874_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_2885_split_sizes_0 = const()[name = string("op_2885_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2885_axis_0 = const()[name = string("op_2885_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2885_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2885_cast_fp16_1 = split(axis = var_2885_axis_0, split_sizes = var_2885_split_sizes_0, x = out_31_cast_fp16)[name = string("op_2885_cast_fp16")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_7_mlp_gate_proj_weight_cast_fp16, x = var_2885_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2902_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_2902_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407383936)))];
            tensor<int32, [2]> var_2908_strides_0 = const()[name = string("op_2908_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2908_pad_type_0 = const()[name = string("op_2908_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2908_pad_0 = const()[name = string("op_2908_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2908_dilations_0 = const()[name = string("op_2908_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2908_groups_0 = const()[name = string("op_2908_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2908_cast_fp16 = conv(dilations = var_2908_dilations_0, groups = var_2908_groups_0, pad = var_2908_pad_0, pad_type = var_2908_pad_type_0, strides = var_2908_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_2885_cast_fp16_0)[name = string("op_2908_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_79_cast_fp16 = mul(x = var_2902_cast_fp16, y = var_2908_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432549824)))];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_77_pad_0 = const()[name = string("hidden_states_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_77_dilations_0 = const()[name = string("hidden_states_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_77_groups_0 = const()[name = string("hidden_states_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_77_cast_fp16 = conv(dilations = hidden_states_77_dilations_0, groups = hidden_states_77_groups_0, pad = hidden_states_77_pad_0, pad_type = hidden_states_77_pad_type_0, strides = hidden_states_77_strides_0, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_79_cast_fp16 = add(x = hidden_states_75_cast_fp16, y = hidden_states_77_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_2926_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2926_cast_fp16")];
            int32 var_2924 = const()[name = string("op_2924"), val = int32(1)];
            bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_65_cast_fp16 = concat(axis = var_2924, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_2926_cast_fp16))[name = string("doubled_65_cast_fp16")];
            tensor<int32, [1]> out_33_axes_0 = const()[name = string("out_33_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_33_gamma_0_to_fp16 = const()[name = string("out_33_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457715712)))];
            fp16 var_2936_to_fp16 = const()[name = string("op_2936_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_2936_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_2947_split_sizes_0 = const()[name = string("op_2947_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2947_axis_0 = const()[name = string("op_2947_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2947_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2947_cast_fp16_1 = split(axis = var_2947_axis_0, split_sizes = var_2947_split_sizes_0, x = out_33_cast_fp16)[name = string("op_2947_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457723968)))];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_49_pad_0 = const()[name = string("query_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_49_dilations_0 = const()[name = string("query_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_49_groups_0 = const()[name = string("query_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_49_cast_fp16 = conv(dilations = query_states_49_dilations_0, groups = query_states_49_groups_0, pad = query_states_49_pad_0, pad_type = query_states_49_pad_type_0, strides = query_states_49_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466112640)))];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_81_pad_0 = const()[name = string("key_states_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_81_dilations_0 = const()[name = string("key_states_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_81_groups_0 = const()[name = string("key_states_81_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_81_cast_fp16 = conv(dilations = key_states_81_dilations_0, groups = key_states_81_groups_0, pad = key_states_81_pad_0, pad_type = key_states_81_pad_type_0, strides = key_states_81_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467161280)))];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_49_pad_0 = const()[name = string("value_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_49_dilations_0 = const()[name = string("value_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_49_groups_0 = const()[name = string("value_states_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_49_cast_fp16 = conv(dilations = value_states_49_dilations_0, groups = value_states_49_groups_0, pad = value_states_49_pad_0, pad_type = value_states_49_pad_type_0, strides = value_states_49_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = var_2947_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_81_cast_fp16 = reshape(shape = concat_96x, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3004_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3004_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3011_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3011_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3015_cast_fp16 = mul(x = x_81_cast_fp16, y = var_308_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<int32, [2]> var_3016_split_sizes_0 = const()[name = string("op_3016_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3016_axis_0 = const()[name = string("op_3016_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3016_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3016_cast_fp16_1 = split(axis = var_3016_axis_0, split_sizes = var_3016_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3016_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_3018_cast_fp16 = mul(x = var_3016_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3018_cast_fp16")];
            int32 var_3020 = const()[name = string("op_3020"), val = int32(-2)];
            bool var_3021_interleave_0 = const()[name = string("op_3021_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3021_cast_fp16 = concat(axis = var_3020, interleave = var_3021_interleave_0, values = (var_3018_cast_fp16, var_3016_cast_fp16_0))[name = string("op_3021_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3022_cast_fp16 = mul(x = var_3021_cast_fp16, y = var_315_cast_fp16)[name = string("op_3022_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_51_cast_fp16 = add(x = var_3015_cast_fp16, y = var_3022_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3028_cast_fp16 = mul(x = var_3004_cast_fp16, y = var_308_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3029_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = var_3004_cast_fp16)[name = string("op_3029_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_3031_cast_fp16 = mul(x = var_3029_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3031_cast_fp16")];
            int32 var_3033 = const()[name = string("op_3033"), val = int32(-2)];
            bool var_3034_interleave_0 = const()[name = string("op_3034_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3034_cast_fp16 = concat(axis = var_3033, interleave = var_3034_interleave_0, values = (var_3031_cast_fp16, var_3029_cast_fp16_0))[name = string("op_3034_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3035_cast_fp16 = mul(x = var_3034_cast_fp16, y = var_315_cast_fp16)[name = string("op_3035_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_85_cast_fp16 = add(x = var_3028_cast_fp16, y = var_3035_cast_fp16)[name = string("key_states_85_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (expand_dims_96, expand_dims_97, position_id, expand_dims_99))[name = string("concat_101")];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> concat_102_values1_0 = const()[name = string("concat_102_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_102_values3_0 = const()[name = string("concat_102_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_102_axis_0 = const()[name = string("concat_102_axis_0"), val = int32(0)];
            bool concat_102_interleave_0 = const()[name = string("concat_102_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_102 = concat(axis = concat_102_axis_0, interleave = concat_102_interleave_0, values = (expand_dims_100, concat_102_values1_0, cache_position_end, concat_102_values3_0))[name = string("concat_102")];
            tensor<int32, [4]> key_states_87_perm_0 = const()[name = string("key_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_87_cast_fp16 = transpose(perm = key_states_87_perm_0, x = key_states_85_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [9, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = key_states_87_cast_fp16, x = coreml_update_state_14)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_16_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_16 = read_state(input = key_cache)[name = string("coreml_update_state_16")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = string("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_3011_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [9, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_101, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_102, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = value_states_51_cast_fp16, x = coreml_update_state_15)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_17_write_state")];
            tensor<fp16, [9, 2, 2048, 128]> coreml_update_state_17 = read_state(input = value_cache)[name = string("coreml_update_state_17")];
            tensor<int32, [4]> var_3105_begin_0 = const()[name = string("op_3105_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3105_end_0 = const()[name = string("op_3105_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3105_end_mask_0 = const()[name = string("op_3105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3105_cast_fp16 = slice_by_index(begin = var_3105_begin_0, end = var_3105_end_0, end_mask = var_3105_end_mask_0, x = coreml_update_state_16)[name = string("op_3105_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3108_axis_0 = const()[name = string("op_3108_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3108_cast_fp16_1 = split(axis = var_3108_axis_0, split_sizes = tile_16, x = var_3105_cast_fp16)[name = string("op_3108_cast_fp16")];
            tensor<int32, [4]> var_3115_begin_0 = const()[name = string("op_3115_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3115_end_0 = const()[name = string("op_3115_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_3115_end_mask_0 = const()[name = string("op_3115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3115_cast_fp16 = slice_by_index(begin = var_3115_begin_0, end = var_3115_end_0, end_mask = var_3115_end_mask_0, x = coreml_update_state_17)[name = string("op_3115_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3118_axis_0 = const()[name = string("op_3118_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3118_cast_fp16_1 = split(axis = var_3118_axis_0, split_sizes = tile_17, x = var_3115_cast_fp16)[name = string("op_3118_cast_fp16")];
            tensor<int32, [2]> var_3121_split_sizes_0 = const()[name = string("op_3121_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3121_axis_0 = const()[name = string("op_3121_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3121_0, tensor<fp16, [1, 8, 128, 1]> var_3121_1 = split(axis = var_3121_axis_0, split_sizes = var_3121_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3121")];
            bool attn_weights_129_transpose_x_0 = const()[name = string("attn_weights_129_transpose_x_0"), val = bool(false)];
            bool attn_weights_129_transpose_y_0 = const()[name = string("attn_weights_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_129_cast_fp16 = matmul(transpose_x = attn_weights_129_transpose_x_0, transpose_y = attn_weights_129_transpose_y_0, x = var_3108_cast_fp16_0, y = var_3121_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3124_to_fp16 = const()[name = string("op_3124_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_131_cast_fp16 = mul(x = attn_weights_129_cast_fp16, y = var_3124_to_fp16)[name = string("attn_weights_131_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_133_cast_fp16 = add(x = attn_weights_131_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_133_cast_fp16")];
            int32 var_3128 = const()[name = string("op_3128"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_135_cast_fp16 = softmax(axis = var_3128, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3134_transpose_x_1 = const()[name = string("op_3134_transpose_x_1"), val = bool(true)];
            bool var_3134_transpose_y_1 = const()[name = string("op_3134_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3134_cast_fp16 = matmul(transpose_x = var_3134_transpose_x_1, transpose_y = var_3134_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3118_cast_fp16_0)[name = string("op_3134_cast_fp16")];
            bool attn_weights_137_transpose_x_0 = const()[name = string("attn_weights_137_transpose_x_0"), val = bool(false)];
            bool attn_weights_137_transpose_y_0 = const()[name = string("attn_weights_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_137_cast_fp16 = matmul(transpose_x = attn_weights_137_transpose_x_0, transpose_y = attn_weights_137_transpose_y_0, x = var_3108_cast_fp16_1, y = var_3121_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3136_to_fp16 = const()[name = string("op_3136_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_139_cast_fp16 = mul(x = attn_weights_137_cast_fp16, y = var_3136_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            int32 var_3140 = const()[name = string("op_3140"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_cast_fp16 = softmax(axis = var_3140, x = attn_weights_141_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_65_transpose_x_1 = const()[name = string("attn_output_65_transpose_x_1"), val = bool(true)];
            bool attn_output_65_transpose_y_1 = const()[name = string("attn_output_65_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_1, transpose_y = attn_output_65_transpose_y_1, x = attn_weights_cast_fp16, y = var_3118_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3148 = const()[name = string("op_3148"), val = int32(1)];
            bool attn_output_67_interleave_0 = const()[name = string("attn_output_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_67_cast_fp16 = concat(axis = var_3148, interleave = attn_output_67_interleave_0, values = (var_3134_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3152_perm_0 = const()[name = string("op_3152_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_107x = const()[name = string("concat_107x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_3152_cast_fp16 = transpose(perm = var_3152_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_cast_fp16 = reshape(shape = concat_107x, x = var_3152_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468209920)))];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_83_pad_0 = const()[name = string("hidden_states_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_83_dilations_0 = const()[name = string("hidden_states_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_83_groups_0 = const()[name = string("hidden_states_83_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_83_cast_fp16 = conv(dilations = hidden_states_83_dilations_0, groups = hidden_states_83_groups_0, pad = hidden_states_83_pad_0, pad_type = hidden_states_83_pad_type_0, strides = hidden_states_83_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_85_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3185_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3185_cast_fp16")];
            int32 var_3183 = const()[name = string("op_3183"), val = int32(1)];
            bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_69_cast_fp16 = concat(axis = var_3183, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3185_cast_fp16))[name = string("doubled_69_cast_fp16")];
            tensor<int32, [1]> out_35_axes_0 = const()[name = string("out_35_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_35_gamma_0_to_fp16 = const()[name = string("out_35_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476598592)))];
            fp16 var_3195_to_fp16 = const()[name = string("op_3195_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_3195_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3206_split_sizes_0 = const()[name = string("op_3206_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3206_axis_0 = const()[name = string("op_3206_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3206_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3206_cast_fp16_1 = split(axis = var_3206_axis_0, split_sizes = var_3206_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3206_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476606848)))];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_8_mlp_gate_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3223_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_3223_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501772736)))];
            tensor<int32, [2]> var_3229_strides_0 = const()[name = string("op_3229_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3229_pad_type_0 = const()[name = string("op_3229_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3229_pad_0 = const()[name = string("op_3229_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3229_dilations_0 = const()[name = string("op_3229_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3229_groups_0 = const()[name = string("op_3229_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3229_cast_fp16 = conv(dilations = var_3229_dilations_0, groups = var_3229_groups_0, pad = var_3229_pad_0, pad_type = var_3229_pad_type_0, strides = var_3229_strides_0, weight = layers_8_mlp_up_proj_weight_to_fp16, x = var_3206_cast_fp16_0)[name = string("op_3229_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_cast_fp16 = mul(x = var_3223_cast_fp16, y = var_3229_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526938624)))];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_87_pad_0 = const()[name = string("hidden_states_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_87_dilations_0 = const()[name = string("hidden_states_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_87_groups_0 = const()[name = string("hidden_states_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_87_cast_fp16 = conv(dilations = hidden_states_87_dilations_0, groups = hidden_states_87_groups_0, pad = hidden_states_87_pad_0, pad_type = hidden_states_87_pad_type_0, strides = hidden_states_87_strides_0, weight = layers_8_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3247_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3247_cast_fp16")];
            int32 var_3245 = const()[name = string("op_3245"), val = int32(1)];
            bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_73_cast_fp16 = concat(axis = var_3245, interleave = doubled_73_interleave_0, values = (hidden_states_cast_fp16, var_3247_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_axes_0 = const()[name = string("out_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_gamma_0_to_fp16 = const()[name = string("out_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552104512)))];
            fp16 var_3257_to_fp16 = const()[name = string("op_3257_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3257_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_3268_split_sizes_0 = const()[name = string("op_3268_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3268_axis_0 = const()[name = string("op_3268_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states, tensor<fp16, [1, 2048, 1, 1]> var_3268_cast_fp16_1 = split(axis = var_3268_axis_0, split_sizes = var_3268_split_sizes_0, x = out_cast_fp16)[name = string("op_3268_cast_fp16")];
        } -> (hidden_states);
}