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, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [1]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_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(64))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524992))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524416))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_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(525312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13120640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13108288))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_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(13126848))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651776))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651200))))[name = string("layers_2_self_attn_v_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(13652096))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26247424))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26235072))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_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(26253632))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26778560))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26777984))))[name = string("layers_3_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_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(26778880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30977408))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30973248))))[name = string("layers_3_self_attn_o_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(30979520))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43566656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43562496))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_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(43568768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093120))))[name = string("layers_4_self_attn_v_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(44094016))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48292544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48288384))))[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(48294656))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60889984))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60877632))))[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(60896192))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73491520))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73479168))))[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(73497728))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86084864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86080704))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_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(86086976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611328))))[name = string("layers_5_self_attn_v_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(86612224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90810752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90806592))))[name = string("layers_5_self_attn_o_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(90812864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103395840))))[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(103414400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997376))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_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(116003648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528000))))[name = string("layers_6_self_attn_v_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(116528896))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120727424))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120723264))))[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(120729536))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133324864))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133312512))))[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(133331072))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145926400))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145914048))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_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(145932608))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158519744))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158515584))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_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(158521856))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046784))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046208))))[name = string("layers_7_self_attn_v_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(159047104))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163245632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241472))))[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(163247744))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175843072))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175830720))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_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(175849280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176374208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176373632))))[name = string("layers_8_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_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(176374528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180573056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180568896))))[name = string("layers_8_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(180575168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193170496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193158144))))[name = string("layers_8_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(193176704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205772032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205759680))))[name = string("layers_8_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_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(205778240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218365376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218361216))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_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(218367488))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218892416))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218891840))))[name = string("layers_9_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_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(218892736))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223091264))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223087104))))[name = string("layers_9_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(223093376))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235688704))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235676352))))[name = string("layers_9_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(235694912))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248290240))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248277888))))[name = string("layers_9_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_9_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(248296448))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260883584))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260879424))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_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(260885696))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410624))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410048))))[name = string("layers_10_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_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(261410944))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265609472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265605312))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(265611584))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278206912))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278194560))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(278213120))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290808448))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290796096))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_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(290814656))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303401792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303397632))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(303403904))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307602432))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307598272))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_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(307604544))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308129472))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308128896))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_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(308129792))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654720))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654144))))[name = string("layers_11_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(308655040))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312853568))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312849408))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(312855680))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325451008))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325438656))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(325457216))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338052544))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040192))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_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(338058752))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350645888))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350641728))))[name = string("layers_11_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(350648000))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354846528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354842368))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_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(354848640))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355373568))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355372992))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_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(355373888))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898816))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898240))))[name = string("layers_12_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(355899136))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360097664))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360093504))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(360099776))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372695104))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372682752))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(372701312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385296640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385284288))))[name = string("layers_12_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_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(385302848))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397889984))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397885824))))[name = string("layers_12_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(397892096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402090624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402086464))))[name = string("layers_13_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_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(402092736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617088))))[name = string("layers_13_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_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(402617984))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142912))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142336))))[name = string("layers_13_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(403143232))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407341760))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407337600))))[name = string("layers_13_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_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(407343872))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419939200))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419926848))))[name = string("layers_13_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_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(419945408))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432532544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432528384))))[name = string("layers_13_mlp_down_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_424 = const()[name = string("op_424"), val = int32(0)];
            bool var_426_exclusive_0 = const()[name = string("op_426_exclusive_0"), val = bool(false)];
            bool var_426_reverse_0 = const()[name = string("op_426_reverse_0"), val = bool(false)];
            tensor<fp16, [1]> var_426_cast_fp16 = cumsum(axis = var_424, exclusive = var_426_exclusive_0, reverse = var_426_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_426_cast_fp16")];
            fp16 var_428_promoted_to_fp16 = const()[name = string("op_428_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> position_offsets_cast_fp16 = sub(x = var_426_cast_fp16, y = var_428_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_431_axes_0 = const()[name = string("op_431_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1]> var_431_cast_fp16 = expand_dims(axes = var_431_axes_0, x = position_offsets_cast_fp16)[name = string("op_431_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_431_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(432534656)))];
            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(440923328)))];
            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_450_perm_0 = const()[name = string("op_450_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_452_axes_0 = const()[name = string("op_452_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_450_cast_fp16 = transpose(perm = var_450_perm_0, x = cos_1_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 1, 128, 1]> var_452_cast_fp16 = expand_dims(axes = var_452_axes_0, x = var_450_cast_fp16)[name = string("op_452_cast_fp16")];
            tensor<int32, [3]> var_457_perm_0 = const()[name = string("op_457_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_459_axes_0 = const()[name = string("op_459_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_457_cast_fp16 = transpose(perm = var_457_perm_0, x = sin_1_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 1, 128, 1]> var_459_cast_fp16 = expand_dims(axes = var_459_axes_0, x = var_457_cast_fp16)[name = string("op_459_cast_fp16")];
            tensor<int32, [1]> var_478_axes_0 = const()[name = string("op_478_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 1, 1]> var_478 = expand_dims(axes = var_478_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_478")];
            tensor<int32, [1, 1, 2048]> var_471 = const()[name = string("op_471"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449312000)))];
            tensor<bool, [1, 1, 2048]> var_479 = greater(x = var_471, y = var_478)[name = string("op_479")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_486_axes_0 = const()[name = string("op_486_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 2048]> var_479_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_479)[name = string("cast_5")];
            tensor<fp16, [1, 1, 1, 2048]> var_486_cast_fp16 = expand_dims(axes = var_486_axes_0, x = var_479_to_fp16)[name = string("op_486_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_490_promoted_to_fp16 = const()[name = string("op_490_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 1]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_486_cast_fp16)[name = string("transpose_87")];
            tensor<bool, [1, 1, 2048, 1]> var_491_cast_fp16 = equal(x = mask_cast_fp16, y = var_490_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            fp16 var_492_to_fp16 = const()[name = string("op_492_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 1]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_492_to_fp16, cond = var_491_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_502_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_502_cast_fp16")];
            int32 var_500 = const()[name = string("op_500"), 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_500, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_502_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(449320256)))];
            fp16 var_512_to_fp16 = const()[name = string("op_512_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_512_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_523_split_sizes_0 = const()[name = string("op_523_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_523_axis_0 = const()[name = string("op_523_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_523_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_523_cast_fp16_1 = split(axis = var_523_axis_0, split_sizes = var_523_split_sizes_0, x = out_1_cast_fp16)[name = string("op_523_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449328512)))];
            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_to_fp16, x = var_523_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457717184)))];
            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_to_fp16, x = var_523_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(458765824)))];
            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_523_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_580_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_580_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_587_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_587_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_591_cast_fp16 = mul(x = x_1_cast_fp16, y = var_452_cast_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [2]> var_592_split_sizes_0 = const()[name = string("op_592_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_592_axis_0 = const()[name = string("op_592_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_592_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_592_cast_fp16_1 = split(axis = var_592_axis_0, split_sizes = var_592_split_sizes_0, x = x_1_cast_fp16)[name = string("op_592_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_594_cast_fp16 = mul(x = var_592_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_594_cast_fp16")];
            int32 var_596 = const()[name = string("op_596"), val = int32(-2)];
            bool var_597_interleave_0 = const()[name = string("op_597_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_597_cast_fp16 = concat(axis = var_596, interleave = var_597_interleave_0, values = (var_594_cast_fp16, var_592_cast_fp16_0))[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_598_cast_fp16 = mul(x = var_597_cast_fp16, y = var_459_cast_fp16)[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_3_cast_fp16 = add(x = var_591_cast_fp16, y = var_598_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_604_cast_fp16 = mul(x = var_580_cast_fp16, y = var_452_cast_fp16)[name = string("op_604_cast_fp16")];
            tensor<int32, [2]> var_605_split_sizes_0 = const()[name = string("op_605_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_605_axis_0 = const()[name = string("op_605_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_605_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_605_cast_fp16_1 = split(axis = var_605_axis_0, split_sizes = var_605_split_sizes_0, x = var_580_cast_fp16)[name = string("op_605_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_607_cast_fp16 = mul(x = var_605_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_607_cast_fp16")];
            int32 var_609 = const()[name = string("op_609"), val = int32(-2)];
            bool var_610_interleave_0 = const()[name = string("op_610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_610_cast_fp16 = concat(axis = var_609, interleave = var_610_interleave_0, values = (var_607_cast_fp16, var_605_cast_fp16_0))[name = string("op_610_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_611_cast_fp16 = mul(x = var_610_cast_fp16, y = var_459_cast_fp16)[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_5_cast_fp16 = add(x = var_604_cast_fp16, y = var_611_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 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_86")];
            tensor<fp16, [14, 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_28_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_28 = read_state(input = key_cache)[name = string("coreml_update_state_28")];
            tensor<fp16, [14, 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_587_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [14, 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_29_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_29 = read_state(input = value_cache)[name = string("coreml_update_state_29")];
            tensor<int32, [4]> var_681_begin_0 = const()[name = string("op_681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_681_end_0 = const()[name = string("op_681_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_681_end_mask_0 = const()[name = string("op_681_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_681_cast_fp16 = slice_by_index(begin = var_681_begin_0, end = var_681_end_0, end_mask = var_681_end_mask_0, x = coreml_update_state_28)[name = string("op_681_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_684_axis_0 = const()[name = string("op_684_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_1 = split(axis = var_684_axis_0, split_sizes = tile_0, x = var_681_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [4]> var_691_begin_0 = const()[name = string("op_691_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_691_end_0 = const()[name = string("op_691_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_691_end_mask_0 = const()[name = string("op_691_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = coreml_update_state_29)[name = string("op_691_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_694_axis_0 = const()[name = string("op_694_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_1 = split(axis = var_694_axis_0, split_sizes = tile_1, x = var_691_cast_fp16)[name = string("op_694_cast_fp16")];
            tensor<int32, [2]> var_697_split_sizes_0 = const()[name = string("op_697_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_697_axis_0 = const()[name = string("op_697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_697_0, tensor<fp16, [1, 8, 128, 1]> var_697_1 = split(axis = var_697_axis_0, split_sizes = var_697_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_697")];
            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_684_cast_fp16_0, y = var_697_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_700_to_fp16 = const()[name = string("op_700_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_700_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_704 = const()[name = string("op_704"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_7_cast_fp16 = softmax(axis = var_704, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_710_transpose_x_1 = const()[name = string("op_710_transpose_x_1"), val = bool(true)];
            bool var_710_transpose_y_1 = const()[name = string("op_710_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_710_cast_fp16 = matmul(transpose_x = var_710_transpose_x_1, transpose_y = var_710_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_694_cast_fp16_0)[name = string("op_710_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_684_cast_fp16_1, y = var_697_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_712_to_fp16 = const()[name = string("op_712_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_712_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_716 = const()[name = string("op_716"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_15_cast_fp16 = softmax(axis = var_716, 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_694_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_724 = const()[name = string("op_724"), 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_724, interleave = attn_output_3_interleave_0, values = (var_710_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_728_perm_0 = const()[name = string("op_728_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_728_cast_fp16 = transpose(perm = var_728_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_728_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(459814464)))];
            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_761_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_761_cast_fp16")];
            int32 var_759 = const()[name = string("op_759"), 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_759, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_761_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(468203136)))];
            fp16 var_771_to_fp16 = const()[name = string("op_771_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_771_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_782_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = out_3_cast_fp16)[name = string("op_782_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468211392)))];
            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_to_fp16, x = var_782_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_799_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_799_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493377280)))];
            tensor<int32, [2]> var_805_strides_0 = const()[name = string("op_805_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_805_pad_type_0 = const()[name = string("op_805_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_805_pad_0 = const()[name = string("op_805_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_805_dilations_0 = const()[name = string("op_805_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_805_groups_0 = const()[name = string("op_805_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_805_cast_fp16 = conv(dilations = var_805_dilations_0, groups = var_805_groups_0, pad = var_805_pad_0, pad_type = var_805_pad_type_0, strides = var_805_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_782_cast_fp16_0)[name = string("op_805_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_9_cast_fp16 = mul(x = var_799_cast_fp16, y = var_805_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518543168)))];
            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_to_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_823_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_823_cast_fp16")];
            int32 var_821 = const()[name = string("op_821"), 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_821, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_823_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(543709056)))];
            fp16 var_833_to_fp16 = const()[name = string("op_833_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_833_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_844_split_sizes_0 = const()[name = string("op_844_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_844_axis_0 = const()[name = string("op_844_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_844_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_844_cast_fp16_1 = split(axis = var_844_axis_0, split_sizes = var_844_split_sizes_0, x = out_5_cast_fp16)[name = string("op_844_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543717312)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552105984)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            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_cast_fp16, x = var_844_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_901_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_901_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_908_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_908_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_912_cast_fp16 = mul(x = x_11_cast_fp16, y = var_452_cast_fp16)[name = string("op_912_cast_fp16")];
            tensor<int32, [2]> var_913_split_sizes_0 = const()[name = string("op_913_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_913_axis_0 = const()[name = string("op_913_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_913_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_913_cast_fp16_1 = split(axis = var_913_axis_0, split_sizes = var_913_split_sizes_0, x = x_11_cast_fp16)[name = string("op_913_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_915_cast_fp16 = mul(x = var_913_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_915_cast_fp16")];
            int32 var_917 = const()[name = string("op_917"), val = int32(-2)];
            bool var_918_interleave_0 = const()[name = string("op_918_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_918_cast_fp16 = concat(axis = var_917, interleave = var_918_interleave_0, values = (var_915_cast_fp16, var_913_cast_fp16_0))[name = string("op_918_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_919_cast_fp16 = mul(x = var_918_cast_fp16, y = var_459_cast_fp16)[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_9_cast_fp16 = add(x = var_912_cast_fp16, y = var_919_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_925_cast_fp16 = mul(x = var_901_cast_fp16, y = var_452_cast_fp16)[name = string("op_925_cast_fp16")];
            tensor<int32, [2]> var_926_split_sizes_0 = const()[name = string("op_926_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_926_axis_0 = const()[name = string("op_926_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_926_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_926_cast_fp16_1 = split(axis = var_926_axis_0, split_sizes = var_926_split_sizes_0, x = var_901_cast_fp16)[name = string("op_926_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_928_cast_fp16 = mul(x = var_926_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_928_cast_fp16")];
            int32 var_930 = const()[name = string("op_930"), val = int32(-2)];
            bool var_931_interleave_0 = const()[name = string("op_931_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_931_cast_fp16 = concat(axis = var_930, interleave = var_931_interleave_0, values = (var_928_cast_fp16, var_926_cast_fp16_0))[name = string("op_931_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_932_cast_fp16 = mul(x = var_931_cast_fp16, y = var_459_cast_fp16)[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_15_cast_fp16 = add(x = var_925_cast_fp16, y = var_932_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_83")];
            tensor<fp16, [14, 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_28)[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_30_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_30 = read_state(input = key_cache)[name = string("coreml_update_state_30")];
            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_908_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [14, 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_29)[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_31_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_31 = read_state(input = value_cache)[name = string("coreml_update_state_31")];
            tensor<int32, [4]> var_1002_begin_0 = const()[name = string("op_1002_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1002_end_0 = const()[name = string("op_1002_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1002_end_mask_0 = const()[name = string("op_1002_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1002_cast_fp16 = slice_by_index(begin = var_1002_begin_0, end = var_1002_end_0, end_mask = var_1002_end_mask_0, x = coreml_update_state_30)[name = string("op_1002_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1005_axis_0 = const()[name = string("op_1005_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_1 = split(axis = var_1005_axis_0, split_sizes = tile_2, x = var_1002_cast_fp16)[name = string("op_1005_cast_fp16")];
            tensor<int32, [4]> var_1012_begin_0 = const()[name = string("op_1012_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1012_end_0 = const()[name = string("op_1012_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1012_end_mask_0 = const()[name = string("op_1012_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1012_cast_fp16 = slice_by_index(begin = var_1012_begin_0, end = var_1012_end_0, end_mask = var_1012_end_mask_0, x = coreml_update_state_31)[name = string("op_1012_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1015_axis_0 = const()[name = string("op_1015_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_1 = split(axis = var_1015_axis_0, split_sizes = tile_3, x = var_1012_cast_fp16)[name = string("op_1015_cast_fp16")];
            tensor<int32, [2]> var_1018_split_sizes_0 = const()[name = string("op_1018_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1018_axis_0 = const()[name = string("op_1018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1018_0, tensor<fp16, [1, 8, 128, 1]> var_1018_1 = split(axis = var_1018_axis_0, split_sizes = var_1018_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1018")];
            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_1005_cast_fp16_0, y = var_1018_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1021_to_fp16 = const()[name = string("op_1021_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_1021_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_1025 = const()[name = string("op_1025"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_23_cast_fp16 = softmax(axis = var_1025, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1031_transpose_x_1 = const()[name = string("op_1031_transpose_x_1"), val = bool(true)];
            bool var_1031_transpose_y_1 = const()[name = string("op_1031_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1031_cast_fp16 = matmul(transpose_x = var_1031_transpose_x_1, transpose_y = var_1031_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1015_cast_fp16_0)[name = string("op_1031_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_1005_cast_fp16_1, y = var_1018_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1033_to_fp16 = const()[name = string("op_1033_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_1033_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_1037 = const()[name = string("op_1037"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_31_cast_fp16 = softmax(axis = var_1037, 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_1015_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1045 = const()[name = string("op_1045"), 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_1045, interleave = attn_output_11_interleave_0, values = (var_1031_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1049_perm_0 = const()[name = string("op_1049_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_1049_cast_fp16 = transpose(perm = var_1049_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1049_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553154624)))];
            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_to_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_1082_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1082_cast_fp16")];
            int32 var_1080 = const()[name = string("op_1080"), 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_1080, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1082_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(561543296)))];
            fp16 var_1092_to_fp16 = const()[name = string("op_1092_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_1092_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1103_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1103_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561551552)))];
            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_to_fp16, x = var_1103_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1120_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [2]> var_1126_strides_0 = const()[name = string("op_1126_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1126_pad_type_0 = const()[name = string("op_1126_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1126_pad_0 = const()[name = string("op_1126_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1126_dilations_0 = const()[name = string("op_1126_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1126_groups_0 = const()[name = string("op_1126_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1126_cast_fp16 = conv(dilations = var_1126_dilations_0, groups = var_1126_groups_0, pad = var_1126_pad_0, pad_type = var_1126_pad_type_0, strides = var_1126_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1103_cast_fp16_0)[name = string("op_1126_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_19_cast_fp16 = mul(x = var_1120_cast_fp16, y = var_1126_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586717440)))];
            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_to_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_1144_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1144_cast_fp16")];
            int32 var_1142 = const()[name = string("op_1142"), 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_1142, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1144_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(611883328)))];
            fp16 var_1154_to_fp16 = const()[name = string("op_1154_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_1154_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1165_split_sizes_0 = const()[name = string("op_1165_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1165_axis_0 = const()[name = string("op_1165_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1165_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1165_cast_fp16_1 = split(axis = var_1165_axis_0, split_sizes = var_1165_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1165_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611891584)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620280256)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            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_cast_fp16, x = var_1165_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_1222_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1222_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_1229_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1229_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1233_cast_fp16 = mul(x = x_21_cast_fp16, y = var_452_cast_fp16)[name = string("op_1233_cast_fp16")];
            tensor<int32, [2]> var_1234_split_sizes_0 = const()[name = string("op_1234_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1234_axis_0 = const()[name = string("op_1234_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1234_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1234_cast_fp16_1 = split(axis = var_1234_axis_0, split_sizes = var_1234_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1234_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_1236_cast_fp16 = mul(x = var_1234_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1236_cast_fp16")];
            int32 var_1238 = const()[name = string("op_1238"), val = int32(-2)];
            bool var_1239_interleave_0 = const()[name = string("op_1239_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1239_cast_fp16 = concat(axis = var_1238, interleave = var_1239_interleave_0, values = (var_1236_cast_fp16, var_1234_cast_fp16_0))[name = string("op_1239_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1240_cast_fp16 = mul(x = var_1239_cast_fp16, y = var_459_cast_fp16)[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_15_cast_fp16 = add(x = var_1233_cast_fp16, y = var_1240_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1246_cast_fp16 = mul(x = var_1222_cast_fp16, y = var_452_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<int32, [2]> var_1247_split_sizes_0 = const()[name = string("op_1247_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1247_axis_0 = const()[name = string("op_1247_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1247_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1247_cast_fp16_1 = split(axis = var_1247_axis_0, split_sizes = var_1247_split_sizes_0, x = var_1222_cast_fp16)[name = string("op_1247_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_1249_cast_fp16 = mul(x = var_1247_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1249_cast_fp16")];
            int32 var_1251 = const()[name = string("op_1251"), val = int32(-2)];
            bool var_1252_interleave_0 = const()[name = string("op_1252_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1252_cast_fp16 = concat(axis = var_1251, interleave = var_1252_interleave_0, values = (var_1249_cast_fp16, var_1247_cast_fp16_0))[name = string("op_1252_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1253_cast_fp16 = mul(x = var_1252_cast_fp16, y = var_459_cast_fp16)[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_25_cast_fp16 = add(x = var_1246_cast_fp16, y = var_1253_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_80")];
            tensor<fp16, [14, 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_30)[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_32_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_32 = read_state(input = key_cache)[name = string("coreml_update_state_32")];
            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_1229_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [14, 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_31)[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_33_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_33 = read_state(input = value_cache)[name = string("coreml_update_state_33")];
            tensor<int32, [4]> var_1323_begin_0 = const()[name = string("op_1323_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1323_end_0 = const()[name = string("op_1323_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1323_end_mask_0 = const()[name = string("op_1323_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1323_cast_fp16 = slice_by_index(begin = var_1323_begin_0, end = var_1323_end_0, end_mask = var_1323_end_mask_0, x = coreml_update_state_32)[name = string("op_1323_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1326_axis_0 = const()[name = string("op_1326_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_1 = split(axis = var_1326_axis_0, split_sizes = tile_4, x = var_1323_cast_fp16)[name = string("op_1326_cast_fp16")];
            tensor<int32, [4]> var_1333_begin_0 = const()[name = string("op_1333_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1333_end_0 = const()[name = string("op_1333_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1333_end_mask_0 = const()[name = string("op_1333_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1333_cast_fp16 = slice_by_index(begin = var_1333_begin_0, end = var_1333_end_0, end_mask = var_1333_end_mask_0, x = coreml_update_state_33)[name = string("op_1333_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1336_axis_0 = const()[name = string("op_1336_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_1 = split(axis = var_1336_axis_0, split_sizes = tile_5, x = var_1333_cast_fp16)[name = string("op_1336_cast_fp16")];
            tensor<int32, [2]> var_1339_split_sizes_0 = const()[name = string("op_1339_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1339_axis_0 = const()[name = string("op_1339_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1339_0, tensor<fp16, [1, 8, 128, 1]> var_1339_1 = split(axis = var_1339_axis_0, split_sizes = var_1339_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1339")];
            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_1326_cast_fp16_0, y = var_1339_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1342_to_fp16 = const()[name = string("op_1342_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_1342_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_1346 = const()[name = string("op_1346"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_39_cast_fp16 = softmax(axis = var_1346, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1352_transpose_x_1 = const()[name = string("op_1352_transpose_x_1"), val = bool(true)];
            bool var_1352_transpose_y_1 = const()[name = string("op_1352_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1352_cast_fp16 = matmul(transpose_x = var_1352_transpose_x_1, transpose_y = var_1352_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1336_cast_fp16_0)[name = string("op_1352_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_1326_cast_fp16_1, y = var_1339_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1354_to_fp16 = const()[name = string("op_1354_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_1354_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_1358 = const()[name = string("op_1358"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_47_cast_fp16 = softmax(axis = var_1358, 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_1336_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1366 = const()[name = string("op_1366"), 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_1366, interleave = attn_output_19_interleave_0, values = (var_1352_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1370_perm_0 = const()[name = string("op_1370_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_1370_cast_fp16 = transpose(perm = var_1370_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1370_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621328896)))];
            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_to_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_1403_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1403_cast_fp16")];
            int32 var_1401 = const()[name = string("op_1401"), 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_1401, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1403_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(629717568)))];
            fp16 var_1413_to_fp16 = const()[name = string("op_1413_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_1413_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1424_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1424_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629725824)))];
            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_to_fp16, x = var_1424_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1441_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1441_cast_fp16")];
            tensor<int32, [2]> var_1447_strides_0 = const()[name = string("op_1447_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1447_pad_type_0 = const()[name = string("op_1447_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1447_pad_0 = const()[name = string("op_1447_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1447_dilations_0 = const()[name = string("op_1447_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1447_groups_0 = const()[name = string("op_1447_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1447_cast_fp16 = conv(dilations = var_1447_dilations_0, groups = var_1447_groups_0, pad = var_1447_pad_0, pad_type = var_1447_pad_type_0, strides = var_1447_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1424_cast_fp16_0)[name = string("op_1447_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_29_cast_fp16 = mul(x = var_1441_cast_fp16, y = var_1447_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654891712)))];
            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_to_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_1465_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1465_cast_fp16")];
            int32 var_1463 = const()[name = string("op_1463"), 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_1463, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1465_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(680057600)))];
            fp16 var_1475_to_fp16 = const()[name = string("op_1475_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_1475_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1486_split_sizes_0 = const()[name = string("op_1486_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1486_axis_0 = const()[name = string("op_1486_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1486_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1486_cast_fp16_1 = split(axis = var_1486_axis_0, split_sizes = var_1486_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1486_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680065856)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688454528)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            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_cast_fp16, x = var_1486_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_1543_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1543_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_1550_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1554_cast_fp16 = mul(x = x_31_cast_fp16, y = var_452_cast_fp16)[name = string("op_1554_cast_fp16")];
            tensor<int32, [2]> var_1555_split_sizes_0 = const()[name = string("op_1555_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1555_axis_0 = const()[name = string("op_1555_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1555_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1555_cast_fp16_1 = split(axis = var_1555_axis_0, split_sizes = var_1555_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1555_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_1557_cast_fp16 = mul(x = var_1555_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1557_cast_fp16")];
            int32 var_1559 = const()[name = string("op_1559"), val = int32(-2)];
            bool var_1560_interleave_0 = const()[name = string("op_1560_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1560_cast_fp16 = concat(axis = var_1559, interleave = var_1560_interleave_0, values = (var_1557_cast_fp16, var_1555_cast_fp16_0))[name = string("op_1560_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1561_cast_fp16 = mul(x = var_1560_cast_fp16, y = var_459_cast_fp16)[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_21_cast_fp16 = add(x = var_1554_cast_fp16, y = var_1561_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1567_cast_fp16 = mul(x = var_1543_cast_fp16, y = var_452_cast_fp16)[name = string("op_1567_cast_fp16")];
            tensor<int32, [2]> var_1568_split_sizes_0 = const()[name = string("op_1568_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1568_axis_0 = const()[name = string("op_1568_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1568_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1568_cast_fp16_1 = split(axis = var_1568_axis_0, split_sizes = var_1568_split_sizes_0, x = var_1543_cast_fp16)[name = string("op_1568_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_1570_cast_fp16 = mul(x = var_1568_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1570_cast_fp16")];
            int32 var_1572 = const()[name = string("op_1572"), val = int32(-2)];
            bool var_1573_interleave_0 = const()[name = string("op_1573_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1573_cast_fp16 = concat(axis = var_1572, interleave = var_1573_interleave_0, values = (var_1570_cast_fp16, var_1568_cast_fp16_0))[name = string("op_1573_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1574_cast_fp16 = mul(x = var_1573_cast_fp16, y = var_459_cast_fp16)[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_35_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1574_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_77")];
            tensor<fp16, [14, 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_32)[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_34_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_34 = read_state(input = key_cache)[name = string("coreml_update_state_34")];
            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_1550_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [14, 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_33)[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_35_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_35 = read_state(input = value_cache)[name = string("coreml_update_state_35")];
            tensor<int32, [4]> var_1644_begin_0 = const()[name = string("op_1644_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1644_end_0 = const()[name = string("op_1644_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1644_end_mask_0 = const()[name = string("op_1644_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1644_cast_fp16 = slice_by_index(begin = var_1644_begin_0, end = var_1644_end_0, end_mask = var_1644_end_mask_0, x = coreml_update_state_34)[name = string("op_1644_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1647_axis_0 = const()[name = string("op_1647_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_1 = split(axis = var_1647_axis_0, split_sizes = tile_6, x = var_1644_cast_fp16)[name = string("op_1647_cast_fp16")];
            tensor<int32, [4]> var_1654_begin_0 = const()[name = string("op_1654_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1654_end_0 = const()[name = string("op_1654_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1654_end_mask_0 = const()[name = string("op_1654_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1654_cast_fp16 = slice_by_index(begin = var_1654_begin_0, end = var_1654_end_0, end_mask = var_1654_end_mask_0, x = coreml_update_state_35)[name = string("op_1654_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1657_axis_0 = const()[name = string("op_1657_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_1 = split(axis = var_1657_axis_0, split_sizes = tile_7, x = var_1654_cast_fp16)[name = string("op_1657_cast_fp16")];
            tensor<int32, [2]> var_1660_split_sizes_0 = const()[name = string("op_1660_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1660_axis_0 = const()[name = string("op_1660_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1660_0, tensor<fp16, [1, 8, 128, 1]> var_1660_1 = split(axis = var_1660_axis_0, split_sizes = var_1660_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1660")];
            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_1647_cast_fp16_0, y = var_1660_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1663_to_fp16 = const()[name = string("op_1663_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_1663_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_1667 = const()[name = string("op_1667"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_55_cast_fp16 = softmax(axis = var_1667, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1673_transpose_x_1 = const()[name = string("op_1673_transpose_x_1"), val = bool(true)];
            bool var_1673_transpose_y_1 = const()[name = string("op_1673_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1673_cast_fp16 = matmul(transpose_x = var_1673_transpose_x_1, transpose_y = var_1673_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1657_cast_fp16_0)[name = string("op_1673_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_1647_cast_fp16_1, y = var_1660_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1675_to_fp16 = const()[name = string("op_1675_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_1675_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_1679 = const()[name = string("op_1679"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_63_cast_fp16 = softmax(axis = var_1679, 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_1657_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1687 = const()[name = string("op_1687"), 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_1687, interleave = attn_output_27_interleave_0, values = (var_1673_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1691_perm_0 = const()[name = string("op_1691_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_1691_cast_fp16 = transpose(perm = var_1691_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1691_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            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_cast_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_1724_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1724_cast_fp16")];
            int32 var_1722 = const()[name = string("op_1722"), 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_1722, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1724_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(689503168)))];
            fp16 var_1734_to_fp16 = const()[name = string("op_1734_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_1734_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1745_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1745_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689511424)))];
            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_to_fp16, x = var_1745_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1762_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1762_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(714677312)))];
            tensor<int32, [2]> var_1768_strides_0 = const()[name = string("op_1768_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1768_pad_type_0 = const()[name = string("op_1768_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1768_pad_0 = const()[name = string("op_1768_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1768_dilations_0 = const()[name = string("op_1768_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1768_groups_0 = const()[name = string("op_1768_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1768_cast_fp16 = conv(dilations = var_1768_dilations_0, groups = var_1768_groups_0, pad = var_1768_pad_0, pad_type = var_1768_pad_type_0, strides = var_1768_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1745_cast_fp16_0)[name = string("op_1768_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_39_cast_fp16 = mul(x = var_1762_cast_fp16, y = var_1768_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_1786_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1786_cast_fp16")];
            int32 var_1784 = const()[name = string("op_1784"), 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_1784, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1786_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(739843200)))];
            fp16 var_1796_to_fp16 = const()[name = string("op_1796_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_1796_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1807_split_sizes_0 = const()[name = string("op_1807_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1807_axis_0 = const()[name = string("op_1807_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1807_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1807_cast_fp16_1 = split(axis = var_1807_axis_0, split_sizes = var_1807_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739851456)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(748240128)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            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_cast_fp16, x = var_1807_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_1864_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1864_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_1871_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1871_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1875_cast_fp16 = mul(x = x_41_cast_fp16, y = var_452_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [2]> var_1876_split_sizes_0 = const()[name = string("op_1876_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1876_axis_0 = const()[name = string("op_1876_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1876_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1876_cast_fp16_1 = split(axis = var_1876_axis_0, split_sizes = var_1876_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1876_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_1878_cast_fp16 = mul(x = var_1876_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1878_cast_fp16")];
            int32 var_1880 = const()[name = string("op_1880"), val = int32(-2)];
            bool var_1881_interleave_0 = const()[name = string("op_1881_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1881_cast_fp16 = concat(axis = var_1880, interleave = var_1881_interleave_0, values = (var_1878_cast_fp16, var_1876_cast_fp16_0))[name = string("op_1881_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1882_cast_fp16 = mul(x = var_1881_cast_fp16, y = var_459_cast_fp16)[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_27_cast_fp16 = add(x = var_1875_cast_fp16, y = var_1882_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1888_cast_fp16 = mul(x = var_1864_cast_fp16, y = var_452_cast_fp16)[name = string("op_1888_cast_fp16")];
            tensor<int32, [2]> var_1889_split_sizes_0 = const()[name = string("op_1889_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1889_axis_0 = const()[name = string("op_1889_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1889_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1889_cast_fp16_1 = split(axis = var_1889_axis_0, split_sizes = var_1889_split_sizes_0, x = var_1864_cast_fp16)[name = string("op_1889_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_1891_cast_fp16 = mul(x = var_1889_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1891_cast_fp16")];
            int32 var_1893 = const()[name = string("op_1893"), val = int32(-2)];
            bool var_1894_interleave_0 = const()[name = string("op_1894_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1894_cast_fp16 = concat(axis = var_1893, interleave = var_1894_interleave_0, values = (var_1891_cast_fp16, var_1889_cast_fp16_0))[name = string("op_1894_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1895_cast_fp16 = mul(x = var_1894_cast_fp16, y = var_459_cast_fp16)[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_45_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1895_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_74")];
            tensor<fp16, [14, 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_34)[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_36_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_36 = read_state(input = key_cache)[name = string("coreml_update_state_36")];
            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_1871_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [14, 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_35)[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_37_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_37 = read_state(input = value_cache)[name = string("coreml_update_state_37")];
            tensor<int32, [4]> var_1965_begin_0 = const()[name = string("op_1965_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1965_end_0 = const()[name = string("op_1965_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1965_end_mask_0 = const()[name = string("op_1965_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1965_cast_fp16 = slice_by_index(begin = var_1965_begin_0, end = var_1965_end_0, end_mask = var_1965_end_mask_0, x = coreml_update_state_36)[name = string("op_1965_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1968_axis_0 = const()[name = string("op_1968_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_1 = split(axis = var_1968_axis_0, split_sizes = tile_8, x = var_1965_cast_fp16)[name = string("op_1968_cast_fp16")];
            tensor<int32, [4]> var_1975_begin_0 = const()[name = string("op_1975_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1975_end_0 = const()[name = string("op_1975_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1975_end_mask_0 = const()[name = string("op_1975_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1975_cast_fp16 = slice_by_index(begin = var_1975_begin_0, end = var_1975_end_0, end_mask = var_1975_end_mask_0, x = coreml_update_state_37)[name = string("op_1975_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1978_axis_0 = const()[name = string("op_1978_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_1 = split(axis = var_1978_axis_0, split_sizes = tile_9, x = var_1975_cast_fp16)[name = string("op_1978_cast_fp16")];
            tensor<int32, [2]> var_1981_split_sizes_0 = const()[name = string("op_1981_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1981_axis_0 = const()[name = string("op_1981_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1981_0, tensor<fp16, [1, 8, 128, 1]> var_1981_1 = split(axis = var_1981_axis_0, split_sizes = var_1981_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1981")];
            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_1968_cast_fp16_0, y = var_1981_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1984_to_fp16 = const()[name = string("op_1984_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_1984_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_1988 = const()[name = string("op_1988"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_71_cast_fp16 = softmax(axis = var_1988, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1994_transpose_x_1 = const()[name = string("op_1994_transpose_x_1"), val = bool(true)];
            bool var_1994_transpose_y_1 = const()[name = string("op_1994_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1994_cast_fp16 = matmul(transpose_x = var_1994_transpose_x_1, transpose_y = var_1994_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1978_cast_fp16_0)[name = string("op_1994_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_1968_cast_fp16_1, y = var_1981_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1996_to_fp16 = const()[name = string("op_1996_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_1996_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_2000 = const()[name = string("op_2000"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_79_cast_fp16 = softmax(axis = var_2000, 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_1978_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2008 = const()[name = string("op_2008"), 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_2008, interleave = attn_output_35_interleave_0, values = (var_1994_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2012_perm_0 = const()[name = string("op_2012_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_2012_cast_fp16 = transpose(perm = var_2012_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2012_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_2045_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2045_cast_fp16")];
            int32 var_2043 = const()[name = string("op_2043"), 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_2043, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2045_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(749288768)))];
            fp16 var_2055_to_fp16 = const()[name = string("op_2055_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_2055_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2066_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2066_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_2066_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2083_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2083_cast_fp16")];
            tensor<int32, [2]> var_2089_strides_0 = const()[name = string("op_2089_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2089_pad_type_0 = const()[name = string("op_2089_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2089_pad_0 = const()[name = string("op_2089_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2089_dilations_0 = const()[name = string("op_2089_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2089_groups_0 = const()[name = string("op_2089_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2089_cast_fp16 = conv(dilations = var_2089_dilations_0, groups = var_2089_groups_0, pad = var_2089_pad_0, pad_type = var_2089_pad_type_0, strides = var_2089_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2066_cast_fp16_0)[name = string("op_2089_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_49_cast_fp16 = mul(x = var_2083_cast_fp16, y = var_2089_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_2107_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2107_cast_fp16")];
            int32 var_2105 = const()[name = string("op_2105"), 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_2105, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2107_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(749297024)))];
            fp16 var_2117_to_fp16 = const()[name = string("op_2117_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_2117_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2128_split_sizes_0 = const()[name = string("op_2128_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2128_axis_0 = const()[name = string("op_2128_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2128_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2128_cast_fp16_1 = split(axis = var_2128_axis_0, split_sizes = var_2128_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749305280)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(757693952)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            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_cast_fp16, x = var_2128_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_2185_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2185_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_2192_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2196_cast_fp16 = mul(x = x_51_cast_fp16, y = var_452_cast_fp16)[name = string("op_2196_cast_fp16")];
            tensor<int32, [2]> var_2197_split_sizes_0 = const()[name = string("op_2197_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2197_axis_0 = const()[name = string("op_2197_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2197_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2197_cast_fp16_1 = split(axis = var_2197_axis_0, split_sizes = var_2197_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2197_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_2199_cast_fp16 = mul(x = var_2197_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2199_cast_fp16")];
            int32 var_2201 = const()[name = string("op_2201"), val = int32(-2)];
            bool var_2202_interleave_0 = const()[name = string("op_2202_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2202_cast_fp16 = concat(axis = var_2201, interleave = var_2202_interleave_0, values = (var_2199_cast_fp16, var_2197_cast_fp16_0))[name = string("op_2202_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2203_cast_fp16 = mul(x = var_2202_cast_fp16, y = var_459_cast_fp16)[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_33_cast_fp16 = add(x = var_2196_cast_fp16, y = var_2203_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2209_cast_fp16 = mul(x = var_2185_cast_fp16, y = var_452_cast_fp16)[name = string("op_2209_cast_fp16")];
            tensor<int32, [2]> var_2210_split_sizes_0 = const()[name = string("op_2210_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2210_axis_0 = const()[name = string("op_2210_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2210_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2210_cast_fp16_1 = split(axis = var_2210_axis_0, split_sizes = var_2210_split_sizes_0, x = var_2185_cast_fp16)[name = string("op_2210_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_2212_cast_fp16 = mul(x = var_2210_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2212_cast_fp16")];
            int32 var_2214 = const()[name = string("op_2214"), val = int32(-2)];
            bool var_2215_interleave_0 = const()[name = string("op_2215_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2215_cast_fp16 = concat(axis = var_2214, interleave = var_2215_interleave_0, values = (var_2212_cast_fp16, var_2210_cast_fp16_0))[name = string("op_2215_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2216_cast_fp16 = mul(x = var_2215_cast_fp16, y = var_459_cast_fp16)[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_55_cast_fp16 = add(x = var_2209_cast_fp16, y = var_2216_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_71")];
            tensor<fp16, [14, 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_36)[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_38_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_38 = read_state(input = key_cache)[name = string("coreml_update_state_38")];
            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_2192_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [14, 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_37)[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_39_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_39 = read_state(input = value_cache)[name = string("coreml_update_state_39")];
            tensor<int32, [4]> var_2286_begin_0 = const()[name = string("op_2286_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2286_end_0 = const()[name = string("op_2286_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2286_end_mask_0 = const()[name = string("op_2286_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2286_cast_fp16 = slice_by_index(begin = var_2286_begin_0, end = var_2286_end_0, end_mask = var_2286_end_mask_0, x = coreml_update_state_38)[name = string("op_2286_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2289_axis_0 = const()[name = string("op_2289_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_1 = split(axis = var_2289_axis_0, split_sizes = tile_10, x = var_2286_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> var_2296_begin_0 = const()[name = string("op_2296_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2296_end_0 = const()[name = string("op_2296_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2296_end_mask_0 = const()[name = string("op_2296_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2296_cast_fp16 = slice_by_index(begin = var_2296_begin_0, end = var_2296_end_0, end_mask = var_2296_end_mask_0, x = coreml_update_state_39)[name = string("op_2296_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2299_axis_0 = const()[name = string("op_2299_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_1 = split(axis = var_2299_axis_0, split_sizes = tile_11, x = var_2296_cast_fp16)[name = string("op_2299_cast_fp16")];
            tensor<int32, [2]> var_2302_split_sizes_0 = const()[name = string("op_2302_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2302_axis_0 = const()[name = string("op_2302_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2302_0, tensor<fp16, [1, 8, 128, 1]> var_2302_1 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2302")];
            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_2289_cast_fp16_0, y = var_2302_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2305_to_fp16 = const()[name = string("op_2305_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_2305_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_2309 = const()[name = string("op_2309"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_87_cast_fp16 = softmax(axis = var_2309, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2315_transpose_x_1 = const()[name = string("op_2315_transpose_x_1"), val = bool(true)];
            bool var_2315_transpose_y_1 = const()[name = string("op_2315_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2315_cast_fp16 = matmul(transpose_x = var_2315_transpose_x_1, transpose_y = var_2315_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2299_cast_fp16_0)[name = string("op_2315_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_2289_cast_fp16_1, y = var_2302_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2317_to_fp16 = const()[name = string("op_2317_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_2317_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_2321 = const()[name = string("op_2321"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_95_cast_fp16 = softmax(axis = var_2321, 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_2299_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2329 = const()[name = string("op_2329"), 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_2329, interleave = attn_output_43_interleave_0, values = (var_2315_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2333_perm_0 = const()[name = string("op_2333_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_2333_cast_fp16 = transpose(perm = var_2333_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2333_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_2366_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2366_cast_fp16")];
            int32 var_2364 = const()[name = string("op_2364"), 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_2364, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2366_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(758742592)))];
            fp16 var_2376_to_fp16 = const()[name = string("op_2376_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_2376_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2387_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2387_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(758750848)))];
            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_to_fp16, x = var_2387_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2404_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [2]> var_2410_strides_0 = const()[name = string("op_2410_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2410_pad_type_0 = const()[name = string("op_2410_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2410_pad_0 = const()[name = string("op_2410_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2410_dilations_0 = const()[name = string("op_2410_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2410_groups_0 = const()[name = string("op_2410_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2410_cast_fp16 = conv(dilations = var_2410_dilations_0, groups = var_2410_groups_0, pad = var_2410_pad_0, pad_type = var_2410_pad_type_0, strides = var_2410_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2387_cast_fp16_0)[name = string("op_2410_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_59_cast_fp16 = mul(x = var_2404_cast_fp16, y = var_2410_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_2428_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2428_cast_fp16")];
            int32 var_2426 = const()[name = string("op_2426"), 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_2426, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2428_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(783916736)))];
            fp16 var_2438_to_fp16 = const()[name = string("op_2438_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_2438_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2449_split_sizes_0 = const()[name = string("op_2449_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2449_axis_0 = const()[name = string("op_2449_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2449_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2449_cast_fp16_1 = split(axis = var_2449_axis_0, split_sizes = var_2449_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2449_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(783924992)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792313664)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            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_cast_fp16, x = var_2449_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_2506_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2506_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_2513_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2517_cast_fp16 = mul(x = x_61_cast_fp16, y = var_452_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<int32, [2]> var_2518_split_sizes_0 = const()[name = string("op_2518_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2518_axis_0 = const()[name = string("op_2518_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2518_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2518_cast_fp16_1 = split(axis = var_2518_axis_0, split_sizes = var_2518_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2518_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_2520_cast_fp16 = mul(x = var_2518_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2520_cast_fp16")];
            int32 var_2522 = const()[name = string("op_2522"), val = int32(-2)];
            bool var_2523_interleave_0 = const()[name = string("op_2523_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2523_cast_fp16 = concat(axis = var_2522, interleave = var_2523_interleave_0, values = (var_2520_cast_fp16, var_2518_cast_fp16_0))[name = string("op_2523_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2524_cast_fp16 = mul(x = var_2523_cast_fp16, y = var_459_cast_fp16)[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_39_cast_fp16 = add(x = var_2517_cast_fp16, y = var_2524_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2530_cast_fp16 = mul(x = var_2506_cast_fp16, y = var_452_cast_fp16)[name = string("op_2530_cast_fp16")];
            tensor<int32, [2]> var_2531_split_sizes_0 = const()[name = string("op_2531_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2531_axis_0 = const()[name = string("op_2531_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2531_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2531_cast_fp16_1 = split(axis = var_2531_axis_0, split_sizes = var_2531_split_sizes_0, x = var_2506_cast_fp16)[name = string("op_2531_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_2533_cast_fp16 = mul(x = var_2531_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2533_cast_fp16")];
            int32 var_2535 = const()[name = string("op_2535"), val = int32(-2)];
            bool var_2536_interleave_0 = const()[name = string("op_2536_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2536_cast_fp16 = concat(axis = var_2535, interleave = var_2536_interleave_0, values = (var_2533_cast_fp16, var_2531_cast_fp16_0))[name = string("op_2536_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2537_cast_fp16 = mul(x = var_2536_cast_fp16, y = var_459_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_65_cast_fp16 = add(x = var_2530_cast_fp16, y = var_2537_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_68")];
            tensor<fp16, [14, 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_38)[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_40_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_40 = read_state(input = key_cache)[name = string("coreml_update_state_40")];
            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_2513_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [14, 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_39)[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_41_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_41 = read_state(input = value_cache)[name = string("coreml_update_state_41")];
            tensor<int32, [4]> var_2607_begin_0 = const()[name = string("op_2607_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2607_end_0 = const()[name = string("op_2607_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2607_end_mask_0 = const()[name = string("op_2607_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2607_cast_fp16 = slice_by_index(begin = var_2607_begin_0, end = var_2607_end_0, end_mask = var_2607_end_mask_0, x = coreml_update_state_40)[name = string("op_2607_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2610_axis_0 = const()[name = string("op_2610_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_1 = split(axis = var_2610_axis_0, split_sizes = tile_12, x = var_2607_cast_fp16)[name = string("op_2610_cast_fp16")];
            tensor<int32, [4]> var_2617_begin_0 = const()[name = string("op_2617_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2617_end_0 = const()[name = string("op_2617_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2617_end_mask_0 = const()[name = string("op_2617_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2617_cast_fp16 = slice_by_index(begin = var_2617_begin_0, end = var_2617_end_0, end_mask = var_2617_end_mask_0, x = coreml_update_state_41)[name = string("op_2617_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2620_axis_0 = const()[name = string("op_2620_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_1 = split(axis = var_2620_axis_0, split_sizes = tile_13, x = var_2617_cast_fp16)[name = string("op_2620_cast_fp16")];
            tensor<int32, [2]> var_2623_split_sizes_0 = const()[name = string("op_2623_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2623_axis_0 = const()[name = string("op_2623_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2623_0, tensor<fp16, [1, 8, 128, 1]> var_2623_1 = split(axis = var_2623_axis_0, split_sizes = var_2623_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2623")];
            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_2610_cast_fp16_0, y = var_2623_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2626_to_fp16 = const()[name = string("op_2626_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_2626_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_2630 = const()[name = string("op_2630"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_103_cast_fp16 = softmax(axis = var_2630, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2636_transpose_x_1 = const()[name = string("op_2636_transpose_x_1"), val = bool(true)];
            bool var_2636_transpose_y_1 = const()[name = string("op_2636_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2636_cast_fp16 = matmul(transpose_x = var_2636_transpose_x_1, transpose_y = var_2636_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2620_cast_fp16_0)[name = string("op_2636_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_2610_cast_fp16_1, y = var_2623_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2638_to_fp16 = const()[name = string("op_2638_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_2638_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_2642 = const()[name = string("op_2642"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_111_cast_fp16 = softmax(axis = var_2642, 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_2620_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2650 = const()[name = string("op_2650"), 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_2650, interleave = attn_output_51_interleave_0, values = (var_2636_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2654_perm_0 = const()[name = string("op_2654_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_2654_cast_fp16 = transpose(perm = var_2654_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2654_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_2687_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2687_cast_fp16")];
            int32 var_2685 = const()[name = string("op_2685"), 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_2685, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2687_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(793362304)))];
            fp16 var_2697_to_fp16 = const()[name = string("op_2697_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_2697_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2708_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2708_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_2708_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2725_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2725_cast_fp16")];
            tensor<int32, [2]> var_2731_strides_0 = const()[name = string("op_2731_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2731_pad_type_0 = const()[name = string("op_2731_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2731_pad_0 = const()[name = string("op_2731_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2731_dilations_0 = const()[name = string("op_2731_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2731_groups_0 = const()[name = string("op_2731_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2731_cast_fp16 = conv(dilations = var_2731_dilations_0, groups = var_2731_groups_0, pad = var_2731_pad_0, pad_type = var_2731_pad_type_0, strides = var_2731_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2708_cast_fp16_0)[name = string("op_2731_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_69_cast_fp16 = mul(x = var_2725_cast_fp16, y = var_2731_cast_fp16)[name = string("x_69_cast_fp16")];
            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_cast_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_2749_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2749_cast_fp16")];
            int32 var_2747 = const()[name = string("op_2747"), 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_2747, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2749_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(793370560)))];
            fp16 var_2759_to_fp16 = const()[name = string("op_2759_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_2759_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2770_split_sizes_0 = const()[name = string("op_2770_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2770_axis_0 = const()[name = string("op_2770_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2770_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2770_cast_fp16_1 = split(axis = var_2770_axis_0, split_sizes = var_2770_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2770_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(793378816)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(801767488)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            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_cast_fp16, x = var_2770_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_2827_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2827_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_2834_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2834_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2838_cast_fp16 = mul(x = x_71_cast_fp16, y = var_452_cast_fp16)[name = string("op_2838_cast_fp16")];
            tensor<int32, [2]> var_2839_split_sizes_0 = const()[name = string("op_2839_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2839_axis_0 = const()[name = string("op_2839_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2839_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2839_cast_fp16_1 = split(axis = var_2839_axis_0, split_sizes = var_2839_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2839_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_2841_cast_fp16 = mul(x = var_2839_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2841_cast_fp16")];
            int32 var_2843 = const()[name = string("op_2843"), val = int32(-2)];
            bool var_2844_interleave_0 = const()[name = string("op_2844_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2844_cast_fp16 = concat(axis = var_2843, interleave = var_2844_interleave_0, values = (var_2841_cast_fp16, var_2839_cast_fp16_0))[name = string("op_2844_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2845_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_459_cast_fp16)[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_45_cast_fp16 = add(x = var_2838_cast_fp16, y = var_2845_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2851_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_452_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [2]> var_2852_split_sizes_0 = const()[name = string("op_2852_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2852_axis_0 = const()[name = string("op_2852_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2852_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2852_cast_fp16_1 = split(axis = var_2852_axis_0, split_sizes = var_2852_split_sizes_0, x = var_2827_cast_fp16)[name = string("op_2852_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_2854_cast_fp16 = mul(x = var_2852_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2854_cast_fp16")];
            int32 var_2856 = const()[name = string("op_2856"), val = int32(-2)];
            bool var_2857_interleave_0 = const()[name = string("op_2857_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2857_cast_fp16 = concat(axis = var_2856, interleave = var_2857_interleave_0, values = (var_2854_cast_fp16, var_2852_cast_fp16_0))[name = string("op_2857_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2858_cast_fp16 = mul(x = var_2857_cast_fp16, y = var_459_cast_fp16)[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_75_cast_fp16 = add(x = var_2851_cast_fp16, y = var_2858_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_65")];
            tensor<fp16, [14, 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_40)[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_42_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_42 = read_state(input = key_cache)[name = string("coreml_update_state_42")];
            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_2834_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [14, 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_41)[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_43_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_43 = read_state(input = value_cache)[name = string("coreml_update_state_43")];
            tensor<int32, [4]> var_2928_begin_0 = const()[name = string("op_2928_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2928_end_0 = const()[name = string("op_2928_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2928_end_mask_0 = const()[name = string("op_2928_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2928_cast_fp16 = slice_by_index(begin = var_2928_begin_0, end = var_2928_end_0, end_mask = var_2928_end_mask_0, x = coreml_update_state_42)[name = string("op_2928_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2931_axis_0 = const()[name = string("op_2931_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_1 = split(axis = var_2931_axis_0, split_sizes = tile_14, x = var_2928_cast_fp16)[name = string("op_2931_cast_fp16")];
            tensor<int32, [4]> var_2938_begin_0 = const()[name = string("op_2938_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2938_end_0 = const()[name = string("op_2938_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2938_end_mask_0 = const()[name = string("op_2938_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2938_cast_fp16 = slice_by_index(begin = var_2938_begin_0, end = var_2938_end_0, end_mask = var_2938_end_mask_0, x = coreml_update_state_43)[name = string("op_2938_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2941_axis_0 = const()[name = string("op_2941_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_1 = split(axis = var_2941_axis_0, split_sizes = tile_15, x = var_2938_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<int32, [2]> var_2944_split_sizes_0 = const()[name = string("op_2944_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2944_axis_0 = const()[name = string("op_2944_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2944_0, tensor<fp16, [1, 8, 128, 1]> var_2944_1 = split(axis = var_2944_axis_0, split_sizes = var_2944_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2944")];
            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_2931_cast_fp16_0, y = var_2944_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2947_to_fp16 = const()[name = string("op_2947_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_2947_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_2951 = const()[name = string("op_2951"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_119_cast_fp16 = softmax(axis = var_2951, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2957_transpose_x_1 = const()[name = string("op_2957_transpose_x_1"), val = bool(true)];
            bool var_2957_transpose_y_1 = const()[name = string("op_2957_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2957_cast_fp16 = matmul(transpose_x = var_2957_transpose_x_1, transpose_y = var_2957_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2941_cast_fp16_0)[name = string("op_2957_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_2931_cast_fp16_1, y = var_2944_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2959_to_fp16 = const()[name = string("op_2959_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_2959_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_2963 = const()[name = string("op_2963"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_127_cast_fp16 = softmax(axis = var_2963, 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_2941_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2971 = const()[name = string("op_2971"), 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_2971, interleave = attn_output_59_interleave_0, values = (var_2957_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2975_perm_0 = const()[name = string("op_2975_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_2975_cast_fp16 = transpose(perm = var_2975_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2975_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_3008_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3008_cast_fp16")];
            int32 var_3006 = const()[name = string("op_3006"), 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_3006, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3008_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(802816128)))];
            fp16 var_3018_to_fp16 = const()[name = string("op_3018_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_3018_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3029_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3029_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_3029_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3046_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3046_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(802824384)))];
            tensor<int32, [2]> var_3052_strides_0 = const()[name = string("op_3052_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3052_pad_type_0 = const()[name = string("op_3052_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3052_pad_0 = const()[name = string("op_3052_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3052_dilations_0 = const()[name = string("op_3052_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3052_groups_0 = const()[name = string("op_3052_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3052_cast_fp16 = conv(dilations = var_3052_dilations_0, groups = var_3052_groups_0, pad = var_3052_pad_0, pad_type = var_3052_pad_type_0, strides = var_3052_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_3029_cast_fp16_0)[name = string("op_3052_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_79_cast_fp16 = mul(x = var_3046_cast_fp16, y = var_3052_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(827990272)))];
            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_3070_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3070_cast_fp16")];
            int32 var_3068 = const()[name = string("op_3068"), 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_3068, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3070_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(853156160)))];
            fp16 var_3080_to_fp16 = const()[name = string("op_3080_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_3080_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3091_split_sizes_0 = const()[name = string("op_3091_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3091_axis_0 = const()[name = string("op_3091_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3091_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3091_cast_fp16_1 = split(axis = var_3091_axis_0, split_sizes = var_3091_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3091_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(853164416)))];
            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_3091_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(861553088)))];
            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_3091_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            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_cast_fp16, x = var_3091_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_3148_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3148_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_3155_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3155_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3159_cast_fp16 = mul(x = x_81_cast_fp16, y = var_452_cast_fp16)[name = string("op_3159_cast_fp16")];
            tensor<int32, [2]> var_3160_split_sizes_0 = const()[name = string("op_3160_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3160_axis_0 = const()[name = string("op_3160_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3160_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3160_cast_fp16_1 = split(axis = var_3160_axis_0, split_sizes = var_3160_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3160_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_3162_cast_fp16 = mul(x = var_3160_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3162_cast_fp16")];
            int32 var_3164 = const()[name = string("op_3164"), val = int32(-2)];
            bool var_3165_interleave_0 = const()[name = string("op_3165_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3165_cast_fp16 = concat(axis = var_3164, interleave = var_3165_interleave_0, values = (var_3162_cast_fp16, var_3160_cast_fp16_0))[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3166_cast_fp16 = mul(x = var_3165_cast_fp16, y = var_459_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_51_cast_fp16 = add(x = var_3159_cast_fp16, y = var_3166_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3172_cast_fp16 = mul(x = var_3148_cast_fp16, y = var_452_cast_fp16)[name = string("op_3172_cast_fp16")];
            tensor<int32, [2]> var_3173_split_sizes_0 = const()[name = string("op_3173_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3173_axis_0 = const()[name = string("op_3173_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3173_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3173_cast_fp16_1 = split(axis = var_3173_axis_0, split_sizes = var_3173_split_sizes_0, x = var_3148_cast_fp16)[name = string("op_3173_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_3175_cast_fp16 = mul(x = var_3173_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3175_cast_fp16")];
            int32 var_3177 = const()[name = string("op_3177"), val = int32(-2)];
            bool var_3178_interleave_0 = const()[name = string("op_3178_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3178_cast_fp16 = concat(axis = var_3177, interleave = var_3178_interleave_0, values = (var_3175_cast_fp16, var_3173_cast_fp16_0))[name = string("op_3178_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3179_cast_fp16 = mul(x = var_3178_cast_fp16, y = var_459_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_85_cast_fp16 = add(x = var_3172_cast_fp16, y = var_3179_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_62")];
            tensor<fp16, [14, 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_42)[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_44_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_44 = read_state(input = key_cache)[name = string("coreml_update_state_44")];
            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_3155_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [14, 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_43)[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_45_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_45 = read_state(input = value_cache)[name = string("coreml_update_state_45")];
            tensor<int32, [4]> var_3249_begin_0 = const()[name = string("op_3249_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3249_end_0 = const()[name = string("op_3249_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3249_end_mask_0 = const()[name = string("op_3249_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3249_cast_fp16 = slice_by_index(begin = var_3249_begin_0, end = var_3249_end_0, end_mask = var_3249_end_mask_0, x = coreml_update_state_44)[name = string("op_3249_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3252_axis_0 = const()[name = string("op_3252_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_1 = split(axis = var_3252_axis_0, split_sizes = tile_16, x = var_3249_cast_fp16)[name = string("op_3252_cast_fp16")];
            tensor<int32, [4]> var_3259_begin_0 = const()[name = string("op_3259_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3259_end_0 = const()[name = string("op_3259_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3259_end_mask_0 = const()[name = string("op_3259_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3259_cast_fp16 = slice_by_index(begin = var_3259_begin_0, end = var_3259_end_0, end_mask = var_3259_end_mask_0, x = coreml_update_state_45)[name = string("op_3259_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3262_axis_0 = const()[name = string("op_3262_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_1 = split(axis = var_3262_axis_0, split_sizes = tile_17, x = var_3259_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<int32, [2]> var_3265_split_sizes_0 = const()[name = string("op_3265_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3265_axis_0 = const()[name = string("op_3265_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3265_0, tensor<fp16, [1, 8, 128, 1]> var_3265_1 = split(axis = var_3265_axis_0, split_sizes = var_3265_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3265")];
            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_3252_cast_fp16_0, y = var_3265_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3268_to_fp16 = const()[name = string("op_3268_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_3268_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_3272 = const()[name = string("op_3272"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_135_cast_fp16 = softmax(axis = var_3272, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3278_transpose_x_1 = const()[name = string("op_3278_transpose_x_1"), val = bool(true)];
            bool var_3278_transpose_y_1 = const()[name = string("op_3278_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3278_cast_fp16 = matmul(transpose_x = var_3278_transpose_x_1, transpose_y = var_3278_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3262_cast_fp16_0)[name = string("op_3278_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_3252_cast_fp16_1, y = var_3265_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3280_to_fp16 = const()[name = string("op_3280_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_3280_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_3284 = const()[name = string("op_3284"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_143_cast_fp16 = softmax(axis = var_3284, x = attn_weights_141_cast_fp16)[name = string("attn_weights_143_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_143_cast_fp16, y = var_3262_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3292 = const()[name = string("op_3292"), 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_3292, interleave = attn_output_67_interleave_0, values = (var_3278_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3296_perm_0 = const()[name = string("op_3296_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_3296_cast_fp16 = transpose(perm = var_3296_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3296_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            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_cast_fp16, x = attn_output_71_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_3329_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3329_cast_fp16")];
            int32 var_3327 = const()[name = string("op_3327"), 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_3327, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3329_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(862601728)))];
            fp16 var_3339_to_fp16 = const()[name = string("op_3339_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_3339_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3350_split_sizes_0 = const()[name = string("op_3350_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3350_axis_0 = const()[name = string("op_3350_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3350_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3350_cast_fp16_1 = split(axis = var_3350_axis_0, split_sizes = var_3350_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3350_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3367_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3367_cast_fp16")];
            tensor<int32, [2]> var_3373_strides_0 = const()[name = string("op_3373_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3373_pad_type_0 = const()[name = string("op_3373_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3373_pad_0 = const()[name = string("op_3373_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3373_dilations_0 = const()[name = string("op_3373_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3373_groups_0 = const()[name = string("op_3373_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3373_cast_fp16 = conv(dilations = var_3373_dilations_0, groups = var_3373_groups_0, pad = var_3373_pad_0, pad_type = var_3373_pad_type_0, strides = var_3373_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("op_3373_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_89_cast_fp16 = mul(x = var_3367_cast_fp16, y = var_3373_cast_fp16)[name = string("x_89_cast_fp16")];
            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_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_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_3391_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3391_cast_fp16")];
            int32 var_3389 = const()[name = string("op_3389"), 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_3389, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3391_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_37_gamma_0_to_fp16 = const()[name = string("out_37_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862609984)))];
            fp16 var_3401_to_fp16 = const()[name = string("op_3401_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_3401_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3412_split_sizes_0 = const()[name = string("op_3412_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3412_axis_0 = const()[name = string("op_3412_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3412_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3412_cast_fp16_1 = split(axis = var_3412_axis_0, split_sizes = var_3412_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3412_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862618240)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(871006912)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3412_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3469_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3469_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3476_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3480_cast_fp16 = mul(x = x_91_cast_fp16, y = var_452_cast_fp16)[name = string("op_3480_cast_fp16")];
            tensor<int32, [2]> var_3481_split_sizes_0 = const()[name = string("op_3481_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3481_axis_0 = const()[name = string("op_3481_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3481_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3481_cast_fp16_1 = split(axis = var_3481_axis_0, split_sizes = var_3481_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3481_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_3483_cast_fp16 = mul(x = var_3481_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3483_cast_fp16")];
            int32 var_3485 = const()[name = string("op_3485"), val = int32(-2)];
            bool var_3486_interleave_0 = const()[name = string("op_3486_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3486_cast_fp16 = concat(axis = var_3485, interleave = var_3486_interleave_0, values = (var_3483_cast_fp16, var_3481_cast_fp16_0))[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3487_cast_fp16 = mul(x = var_3486_cast_fp16, y = var_459_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_57_cast_fp16 = add(x = var_3480_cast_fp16, y = var_3487_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3493_cast_fp16 = mul(x = var_3469_cast_fp16, y = var_452_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [2]> var_3494_split_sizes_0 = const()[name = string("op_3494_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3494_axis_0 = const()[name = string("op_3494_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3494_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3494_cast_fp16_1 = split(axis = var_3494_axis_0, split_sizes = var_3494_split_sizes_0, x = var_3469_cast_fp16)[name = string("op_3494_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_3496_cast_fp16 = mul(x = var_3494_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3496_cast_fp16")];
            int32 var_3498 = const()[name = string("op_3498"), val = int32(-2)];
            bool var_3499_interleave_0 = const()[name = string("op_3499_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3499_cast_fp16 = concat(axis = var_3498, interleave = var_3499_interleave_0, values = (var_3496_cast_fp16, var_3494_cast_fp16_0))[name = string("op_3499_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3500_cast_fp16 = mul(x = var_3499_cast_fp16, y = var_459_cast_fp16)[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_95_cast_fp16 = add(x = var_3493_cast_fp16, y = var_3500_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_44)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_46_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_46 = read_state(input = key_cache)[name = string("coreml_update_state_46")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3476_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_45)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_47_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_47 = read_state(input = value_cache)[name = string("coreml_update_state_47")];
            tensor<int32, [4]> var_3570_begin_0 = const()[name = string("op_3570_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3570_end_0 = const()[name = string("op_3570_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3570_end_mask_0 = const()[name = string("op_3570_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = coreml_update_state_46)[name = string("op_3570_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3573_axis_0 = const()[name = string("op_3573_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_1 = split(axis = var_3573_axis_0, split_sizes = tile_18, x = var_3570_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [4]> var_3580_begin_0 = const()[name = string("op_3580_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3580_end_0 = const()[name = string("op_3580_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3580_end_mask_0 = const()[name = string("op_3580_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3580_cast_fp16 = slice_by_index(begin = var_3580_begin_0, end = var_3580_end_0, end_mask = var_3580_end_mask_0, x = coreml_update_state_47)[name = string("op_3580_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3583_axis_0 = const()[name = string("op_3583_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_1 = split(axis = var_3583_axis_0, split_sizes = tile_19, x = var_3580_cast_fp16)[name = string("op_3583_cast_fp16")];
            tensor<int32, [2]> var_3586_split_sizes_0 = const()[name = string("op_3586_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3586_axis_0 = const()[name = string("op_3586_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3586_0, tensor<fp16, [1, 8, 128, 1]> var_3586_1 = split(axis = var_3586_axis_0, split_sizes = var_3586_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3586")];
            bool attn_weights_145_transpose_x_0 = const()[name = string("attn_weights_145_transpose_x_0"), val = bool(false)];
            bool attn_weights_145_transpose_y_0 = const()[name = string("attn_weights_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_145_cast_fp16 = matmul(transpose_x = attn_weights_145_transpose_x_0, transpose_y = attn_weights_145_transpose_y_0, x = var_3573_cast_fp16_0, y = var_3586_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3589_to_fp16 = const()[name = string("op_3589_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_147_cast_fp16 = mul(x = attn_weights_145_cast_fp16, y = var_3589_to_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_149_cast_fp16 = add(x = attn_weights_147_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_149_cast_fp16")];
            int32 var_3593 = const()[name = string("op_3593"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_151_cast_fp16 = softmax(axis = var_3593, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3599_transpose_x_1 = const()[name = string("op_3599_transpose_x_1"), val = bool(true)];
            bool var_3599_transpose_y_1 = const()[name = string("op_3599_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3599_cast_fp16 = matmul(transpose_x = var_3599_transpose_x_1, transpose_y = var_3599_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3583_cast_fp16_0)[name = string("op_3599_cast_fp16")];
            bool attn_weights_153_transpose_x_0 = const()[name = string("attn_weights_153_transpose_x_0"), val = bool(false)];
            bool attn_weights_153_transpose_y_0 = const()[name = string("attn_weights_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_153_cast_fp16 = matmul(transpose_x = attn_weights_153_transpose_x_0, transpose_y = attn_weights_153_transpose_y_0, x = var_3573_cast_fp16_1, y = var_3586_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3601_to_fp16 = const()[name = string("op_3601_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_155_cast_fp16 = mul(x = attn_weights_153_cast_fp16, y = var_3601_to_fp16)[name = string("attn_weights_155_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_157_cast_fp16 = add(x = attn_weights_155_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_157_cast_fp16")];
            int32 var_3605 = const()[name = string("op_3605"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_159_cast_fp16 = softmax(axis = var_3605, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3583_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3613 = const()[name = string("op_3613"), val = int32(1)];
            bool attn_output_75_interleave_0 = const()[name = string("attn_output_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_75_cast_fp16 = concat(axis = var_3613, interleave = attn_output_75_interleave_0, values = (var_3599_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3617_perm_0 = const()[name = string("op_3617_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_3617_cast_fp16 = transpose(perm = var_3617_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3617_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3650_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3650_cast_fp16")];
            int32 var_3648 = const()[name = string("op_3648"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_77_cast_fp16 = concat(axis = var_3648, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3650_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872055552)))];
            fp16 var_3660_to_fp16 = const()[name = string("op_3660_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3660_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3671_split_sizes_0 = const()[name = string("op_3671_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3671_axis_0 = const()[name = string("op_3671_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3671_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3671_cast_fp16_1 = split(axis = var_3671_axis_0, split_sizes = var_3671_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3671_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3688_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3688_cast_fp16")];
            tensor<int32, [2]> var_3694_strides_0 = const()[name = string("op_3694_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3694_pad_type_0 = const()[name = string("op_3694_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3694_pad_0 = const()[name = string("op_3694_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3694_dilations_0 = const()[name = string("op_3694_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3694_groups_0 = const()[name = string("op_3694_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3694_cast_fp16 = conv(dilations = var_3694_dilations_0, groups = var_3694_groups_0, pad = var_3694_pad_0, pad_type = var_3694_pad_type_0, strides = var_3694_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("op_3694_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_99_cast_fp16 = mul(x = var_3688_cast_fp16, y = var_3694_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3712_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3712_cast_fp16")];
            int32 var_3710 = const()[name = string("op_3710"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_81_cast_fp16 = concat(axis = var_3710, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3712_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872063808)))];
            fp16 var_3722_to_fp16 = const()[name = string("op_3722_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3722_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3733_split_sizes_0 = const()[name = string("op_3733_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3733_axis_0 = const()[name = string("op_3733_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3733_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3733_cast_fp16_1 = split(axis = var_3733_axis_0, split_sizes = var_3733_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872072064)))];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(880460736)))];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_cast_fp16, x = var_3733_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3790_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3790_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3797_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3797_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3801_cast_fp16 = mul(x = x_101_cast_fp16, y = var_452_cast_fp16)[name = string("op_3801_cast_fp16")];
            tensor<int32, [2]> var_3802_split_sizes_0 = const()[name = string("op_3802_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3802_axis_0 = const()[name = string("op_3802_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3802_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3802_cast_fp16_1 = split(axis = var_3802_axis_0, split_sizes = var_3802_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3802_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_3804_cast_fp16 = mul(x = var_3802_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3804_cast_fp16")];
            int32 var_3806 = const()[name = string("op_3806"), val = int32(-2)];
            bool var_3807_interleave_0 = const()[name = string("op_3807_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3807_cast_fp16 = concat(axis = var_3806, interleave = var_3807_interleave_0, values = (var_3804_cast_fp16, var_3802_cast_fp16_0))[name = string("op_3807_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3808_cast_fp16 = mul(x = var_3807_cast_fp16, y = var_459_cast_fp16)[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_63_cast_fp16 = add(x = var_3801_cast_fp16, y = var_3808_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3814_cast_fp16 = mul(x = var_3790_cast_fp16, y = var_452_cast_fp16)[name = string("op_3814_cast_fp16")];
            tensor<int32, [2]> var_3815_split_sizes_0 = const()[name = string("op_3815_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3815_axis_0 = const()[name = string("op_3815_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3815_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3815_cast_fp16_1 = split(axis = var_3815_axis_0, split_sizes = var_3815_split_sizes_0, x = var_3790_cast_fp16)[name = string("op_3815_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_3817_cast_fp16 = mul(x = var_3815_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3817_cast_fp16")];
            int32 var_3819 = const()[name = string("op_3819"), val = int32(-2)];
            bool var_3820_interleave_0 = const()[name = string("op_3820_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3820_cast_fp16 = concat(axis = var_3819, interleave = var_3820_interleave_0, values = (var_3817_cast_fp16, var_3815_cast_fp16_0))[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3821_cast_fp16 = mul(x = var_3820_cast_fp16, y = var_459_cast_fp16)[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_105_cast_fp16 = add(x = var_3814_cast_fp16, y = var_3821_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_46)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_48_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_48 = read_state(input = key_cache)[name = string("coreml_update_state_48")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3797_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_47)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_49_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_49 = read_state(input = value_cache)[name = string("coreml_update_state_49")];
            tensor<int32, [4]> var_3891_begin_0 = const()[name = string("op_3891_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3891_end_0 = const()[name = string("op_3891_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3891_end_mask_0 = const()[name = string("op_3891_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3891_cast_fp16 = slice_by_index(begin = var_3891_begin_0, end = var_3891_end_0, end_mask = var_3891_end_mask_0, x = coreml_update_state_48)[name = string("op_3891_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3894_axis_0 = const()[name = string("op_3894_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_1 = split(axis = var_3894_axis_0, split_sizes = tile_20, x = var_3891_cast_fp16)[name = string("op_3894_cast_fp16")];
            tensor<int32, [4]> var_3901_begin_0 = const()[name = string("op_3901_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3901_end_0 = const()[name = string("op_3901_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3901_end_mask_0 = const()[name = string("op_3901_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3901_cast_fp16 = slice_by_index(begin = var_3901_begin_0, end = var_3901_end_0, end_mask = var_3901_end_mask_0, x = coreml_update_state_49)[name = string("op_3901_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3904_axis_0 = const()[name = string("op_3904_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_1 = split(axis = var_3904_axis_0, split_sizes = tile_21, x = var_3901_cast_fp16)[name = string("op_3904_cast_fp16")];
            tensor<int32, [2]> var_3907_split_sizes_0 = const()[name = string("op_3907_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3907_axis_0 = const()[name = string("op_3907_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3907_0, tensor<fp16, [1, 8, 128, 1]> var_3907_1 = split(axis = var_3907_axis_0, split_sizes = var_3907_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3907")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3894_cast_fp16_0, y = var_3907_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3910_to_fp16 = const()[name = string("op_3910_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3910_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3914 = const()[name = string("op_3914"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_167_cast_fp16 = softmax(axis = var_3914, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3920_transpose_x_1 = const()[name = string("op_3920_transpose_x_1"), val = bool(true)];
            bool var_3920_transpose_y_1 = const()[name = string("op_3920_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3920_cast_fp16 = matmul(transpose_x = var_3920_transpose_x_1, transpose_y = var_3920_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3904_cast_fp16_0)[name = string("op_3920_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3894_cast_fp16_1, y = var_3907_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3922_to_fp16 = const()[name = string("op_3922_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3922_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3926 = const()[name = string("op_3926"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_175_cast_fp16 = softmax(axis = var_3926, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3904_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3934 = const()[name = string("op_3934"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_83_cast_fp16 = concat(axis = var_3934, interleave = attn_output_83_interleave_0, values = (var_3920_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3938_perm_0 = const()[name = string("op_3938_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_3938_cast_fp16 = transpose(perm = var_3938_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3938_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3971_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3971_cast_fp16")];
            int32 var_3969 = const()[name = string("op_3969"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_85_cast_fp16 = concat(axis = var_3969, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3971_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881509376)))];
            fp16 var_3981_to_fp16 = const()[name = string("op_3981_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3981_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3992_split_sizes_0 = const()[name = string("op_3992_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3992_axis_0 = const()[name = string("op_3992_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3992_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3992_cast_fp16_1 = split(axis = var_3992_axis_0, split_sizes = var_3992_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3992_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4009_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [2]> var_4015_strides_0 = const()[name = string("op_4015_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4015_pad_type_0 = const()[name = string("op_4015_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4015_pad_0 = const()[name = string("op_4015_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4015_dilations_0 = const()[name = string("op_4015_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4015_groups_0 = const()[name = string("op_4015_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4015_cast_fp16 = conv(dilations = var_4015_dilations_0, groups = var_4015_groups_0, pad = var_4015_pad_0, pad_type = var_4015_pad_type_0, strides = var_4015_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("op_4015_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_109_cast_fp16 = mul(x = var_4009_cast_fp16, y = var_4015_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4033_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4033_cast_fp16")];
            int32 var_4031 = const()[name = string("op_4031"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_89_cast_fp16 = concat(axis = var_4031, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4033_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881517632)))];
            fp16 var_4043_to_fp16 = const()[name = string("op_4043_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4043_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4054_split_sizes_0 = const()[name = string("op_4054_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4054_axis_0 = const()[name = string("op_4054_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4054_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4054_cast_fp16_1 = split(axis = var_4054_axis_0, split_sizes = var_4054_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4111_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4111_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4118_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4118_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4122_cast_fp16 = mul(x = x_111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4122_cast_fp16")];
            tensor<int32, [2]> var_4123_split_sizes_0 = const()[name = string("op_4123_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4123_axis_0 = const()[name = string("op_4123_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_4123_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_4123_cast_fp16_1 = split(axis = var_4123_axis_0, split_sizes = var_4123_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4123_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_4125_cast_fp16 = mul(x = var_4123_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4125_cast_fp16")];
            int32 var_4127 = const()[name = string("op_4127"), val = int32(-2)];
            bool var_4128_interleave_0 = const()[name = string("op_4128_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_4128_cast_fp16 = concat(axis = var_4127, interleave = var_4128_interleave_0, values = (var_4125_cast_fp16, var_4123_cast_fp16_0))[name = string("op_4128_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4129_cast_fp16 = mul(x = var_4128_cast_fp16, y = var_459_cast_fp16)[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_69_cast_fp16 = add(x = var_4122_cast_fp16, y = var_4129_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4135_cast_fp16 = mul(x = var_4111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4135_cast_fp16")];
            tensor<int32, [2]> var_4136_split_sizes_0 = const()[name = string("op_4136_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4136_axis_0 = const()[name = string("op_4136_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_4136_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_4136_cast_fp16_1 = split(axis = var_4136_axis_0, split_sizes = var_4136_split_sizes_0, x = var_4111_cast_fp16)[name = string("op_4136_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_4138_cast_fp16 = mul(x = var_4136_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4138_cast_fp16")];
            int32 var_4140 = const()[name = string("op_4140"), val = int32(-2)];
            bool var_4141_interleave_0 = const()[name = string("op_4141_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_4141_cast_fp16 = concat(axis = var_4140, interleave = var_4141_interleave_0, values = (var_4138_cast_fp16, var_4136_cast_fp16_0))[name = string("op_4141_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4142_cast_fp16 = mul(x = var_4141_cast_fp16, y = var_459_cast_fp16)[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_115_cast_fp16 = add(x = var_4135_cast_fp16, y = var_4142_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_48)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_50_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_50 = read_state(input = key_cache)[name = string("coreml_update_state_50")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4118_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_49)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_51_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_51 = read_state(input = value_cache)[name = string("coreml_update_state_51")];
            tensor<int32, [4]> var_4212_begin_0 = const()[name = string("op_4212_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4212_end_0 = const()[name = string("op_4212_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4212_end_mask_0 = const()[name = string("op_4212_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4212_cast_fp16 = slice_by_index(begin = var_4212_begin_0, end = var_4212_end_0, end_mask = var_4212_end_mask_0, x = coreml_update_state_50)[name = string("op_4212_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4215_axis_0 = const()[name = string("op_4215_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_1 = split(axis = var_4215_axis_0, split_sizes = tile_22, x = var_4212_cast_fp16)[name = string("op_4215_cast_fp16")];
            tensor<int32, [4]> var_4222_begin_0 = const()[name = string("op_4222_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4222_end_0 = const()[name = string("op_4222_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4222_end_mask_0 = const()[name = string("op_4222_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4222_cast_fp16 = slice_by_index(begin = var_4222_begin_0, end = var_4222_end_0, end_mask = var_4222_end_mask_0, x = coreml_update_state_51)[name = string("op_4222_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4225_axis_0 = const()[name = string("op_4225_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_1 = split(axis = var_4225_axis_0, split_sizes = tile_23, x = var_4222_cast_fp16)[name = string("op_4225_cast_fp16")];
            tensor<int32, [2]> var_4228_split_sizes_0 = const()[name = string("op_4228_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4228_axis_0 = const()[name = string("op_4228_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_4228_0, tensor<fp16, [1, 8, 128, 1]> var_4228_1 = split(axis = var_4228_axis_0, split_sizes = var_4228_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4228")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4215_cast_fp16_0, y = var_4228_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4231_to_fp16 = const()[name = string("op_4231_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4231_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4235 = const()[name = string("op_4235"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_183_cast_fp16 = softmax(axis = var_4235, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4241_transpose_x_1 = const()[name = string("op_4241_transpose_x_1"), val = bool(true)];
            bool var_4241_transpose_y_1 = const()[name = string("op_4241_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_4241_cast_fp16 = matmul(transpose_x = var_4241_transpose_x_1, transpose_y = var_4241_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4225_cast_fp16_0)[name = string("op_4241_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4215_cast_fp16_1, y = var_4228_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4243_to_fp16 = const()[name = string("op_4243_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4243_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4247 = const()[name = string("op_4247"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_191_cast_fp16 = softmax(axis = var_4247, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4225_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4255 = const()[name = string("op_4255"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_91_cast_fp16 = concat(axis = var_4255, interleave = attn_output_91_interleave_0, values = (var_4241_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4259_perm_0 = const()[name = string("op_4259_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_4259_cast_fp16 = transpose(perm = var_4259_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4259_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4292_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4292_cast_fp16")];
            int32 var_4290 = const()[name = string("op_4290"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_93_cast_fp16 = concat(axis = var_4290, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4292_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881525888)))];
            fp16 var_4302_to_fp16 = const()[name = string("op_4302_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4302_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4313_split_sizes_0 = const()[name = string("op_4313_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4313_axis_0 = const()[name = string("op_4313_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4313_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4313_cast_fp16_1 = split(axis = var_4313_axis_0, split_sizes = var_4313_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4313_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4330_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4330_cast_fp16")];
            tensor<int32, [2]> var_4336_strides_0 = const()[name = string("op_4336_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4336_pad_type_0 = const()[name = string("op_4336_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4336_pad_0 = const()[name = string("op_4336_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4336_dilations_0 = const()[name = string("op_4336_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4336_groups_0 = const()[name = string("op_4336_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4336_cast_fp16 = conv(dilations = var_4336_dilations_0, groups = var_4336_groups_0, pad = var_4336_pad_0, pad_type = var_4336_pad_type_0, strides = var_4336_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("op_4336_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_119_cast_fp16 = mul(x = var_4330_cast_fp16, y = var_4336_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_cast_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4354_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4354_cast_fp16")];
            int32 var_4352 = const()[name = string("op_4352"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_97_cast_fp16 = concat(axis = var_4352, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4354_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881534144)))];
            fp16 var_4364_to_fp16 = const()[name = string("op_4364_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4364_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4375_split_sizes_0 = const()[name = string("op_4375_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4375_axis_0 = const()[name = string("op_4375_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4375_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4375_cast_fp16_1 = split(axis = var_4375_axis_0, split_sizes = var_4375_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4375_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4432_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4432_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4439_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4439_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4443_cast_fp16 = mul(x = x_121_cast_fp16, y = var_452_cast_fp16)[name = string("op_4443_cast_fp16")];
            tensor<int32, [2]> var_4444_split_sizes_0 = const()[name = string("op_4444_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4444_axis_0 = const()[name = string("op_4444_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_4444_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_4444_cast_fp16_1 = split(axis = var_4444_axis_0, split_sizes = var_4444_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4444_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_4446_cast_fp16 = mul(x = var_4444_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4446_cast_fp16")];
            int32 var_4448 = const()[name = string("op_4448"), val = int32(-2)];
            bool var_4449_interleave_0 = const()[name = string("op_4449_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_4449_cast_fp16 = concat(axis = var_4448, interleave = var_4449_interleave_0, values = (var_4446_cast_fp16, var_4444_cast_fp16_0))[name = string("op_4449_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4450_cast_fp16 = mul(x = var_4449_cast_fp16, y = var_459_cast_fp16)[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_75_cast_fp16 = add(x = var_4443_cast_fp16, y = var_4450_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4456_cast_fp16 = mul(x = var_4432_cast_fp16, y = var_452_cast_fp16)[name = string("op_4456_cast_fp16")];
            tensor<int32, [2]> var_4457_split_sizes_0 = const()[name = string("op_4457_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4457_axis_0 = const()[name = string("op_4457_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_4457_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_4457_cast_fp16_1 = split(axis = var_4457_axis_0, split_sizes = var_4457_split_sizes_0, x = var_4432_cast_fp16)[name = string("op_4457_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_4459_cast_fp16 = mul(x = var_4457_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4459_cast_fp16")];
            int32 var_4461 = const()[name = string("op_4461"), val = int32(-2)];
            bool var_4462_interleave_0 = const()[name = string("op_4462_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_4462_cast_fp16 = concat(axis = var_4461, interleave = var_4462_interleave_0, values = (var_4459_cast_fp16, var_4457_cast_fp16_0))[name = string("op_4462_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4463_cast_fp16 = mul(x = var_4462_cast_fp16, y = var_459_cast_fp16)[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_125_cast_fp16 = add(x = var_4456_cast_fp16, y = var_4463_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_50)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_52_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_52 = read_state(input = key_cache)[name = string("coreml_update_state_52")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4439_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_51)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_53_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_53 = read_state(input = value_cache)[name = string("coreml_update_state_53")];
            tensor<int32, [4]> var_4533_begin_0 = const()[name = string("op_4533_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4533_end_0 = const()[name = string("op_4533_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4533_end_mask_0 = const()[name = string("op_4533_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4533_cast_fp16 = slice_by_index(begin = var_4533_begin_0, end = var_4533_end_0, end_mask = var_4533_end_mask_0, x = coreml_update_state_52)[name = string("op_4533_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4536_axis_0 = const()[name = string("op_4536_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_1 = split(axis = var_4536_axis_0, split_sizes = tile_24, x = var_4533_cast_fp16)[name = string("op_4536_cast_fp16")];
            tensor<int32, [4]> var_4543_begin_0 = const()[name = string("op_4543_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4543_end_0 = const()[name = string("op_4543_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4543_end_mask_0 = const()[name = string("op_4543_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4543_cast_fp16 = slice_by_index(begin = var_4543_begin_0, end = var_4543_end_0, end_mask = var_4543_end_mask_0, x = coreml_update_state_53)[name = string("op_4543_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4546_axis_0 = const()[name = string("op_4546_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_1 = split(axis = var_4546_axis_0, split_sizes = tile_25, x = var_4543_cast_fp16)[name = string("op_4546_cast_fp16")];
            tensor<int32, [2]> var_4549_split_sizes_0 = const()[name = string("op_4549_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4549_axis_0 = const()[name = string("op_4549_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_4549_0, tensor<fp16, [1, 8, 128, 1]> var_4549_1 = split(axis = var_4549_axis_0, split_sizes = var_4549_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4549")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4536_cast_fp16_0, y = var_4549_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4552_to_fp16 = const()[name = string("op_4552_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4552_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4556 = const()[name = string("op_4556"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_199_cast_fp16 = softmax(axis = var_4556, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4562_transpose_x_1 = const()[name = string("op_4562_transpose_x_1"), val = bool(true)];
            bool var_4562_transpose_y_1 = const()[name = string("op_4562_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_4562_cast_fp16 = matmul(transpose_x = var_4562_transpose_x_1, transpose_y = var_4562_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4546_cast_fp16_0)[name = string("op_4562_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4536_cast_fp16_1, y = var_4549_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4564_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4568 = const()[name = string("op_4568"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_207_cast_fp16 = softmax(axis = var_4568, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4546_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4576 = const()[name = string("op_4576"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_99_cast_fp16 = concat(axis = var_4576, interleave = attn_output_99_interleave_0, values = (var_4562_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4580_perm_0 = const()[name = string("op_4580_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_4580_cast_fp16 = transpose(perm = var_4580_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4580_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4613_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4613_cast_fp16")];
            int32 var_4611 = const()[name = string("op_4611"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_101_cast_fp16 = concat(axis = var_4611, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4613_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881542400)))];
            fp16 var_4623_to_fp16 = const()[name = string("op_4623_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4623_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4634_split_sizes_0 = const()[name = string("op_4634_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4634_axis_0 = const()[name = string("op_4634_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4634_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4634_cast_fp16_1 = split(axis = var_4634_axis_0, split_sizes = var_4634_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4634_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4651_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4651_cast_fp16")];
            tensor<int32, [2]> var_4657_strides_0 = const()[name = string("op_4657_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4657_pad_type_0 = const()[name = string("op_4657_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4657_pad_0 = const()[name = string("op_4657_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4657_dilations_0 = const()[name = string("op_4657_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4657_groups_0 = const()[name = string("op_4657_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4657_cast_fp16 = conv(dilations = var_4657_dilations_0, groups = var_4657_groups_0, pad = var_4657_pad_0, pad_type = var_4657_pad_type_0, strides = var_4657_strides_0, weight = layers_12_mlp_up_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("op_4657_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_129_cast_fp16 = mul(x = var_4651_cast_fp16, y = var_4657_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_cast_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4675_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4675_cast_fp16")];
            int32 var_4673 = const()[name = string("op_4673"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_105_cast_fp16 = concat(axis = var_4673, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4675_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881550656)))];
            fp16 var_4685_to_fp16 = const()[name = string("op_4685_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4685_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4696_split_sizes_0 = const()[name = string("op_4696_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4696_axis_0 = const()[name = string("op_4696_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4696_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4696_cast_fp16_1 = split(axis = var_4696_axis_0, split_sizes = var_4696_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4696_cast_fp16")];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4753_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4753_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4760_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4760_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4764_cast_fp16 = mul(x = x_131_cast_fp16, y = var_452_cast_fp16)[name = string("op_4764_cast_fp16")];
            tensor<int32, [2]> var_4765_split_sizes_0 = const()[name = string("op_4765_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4765_axis_0 = const()[name = string("op_4765_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_4765_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_4765_cast_fp16_1 = split(axis = var_4765_axis_0, split_sizes = var_4765_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4765_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_4767_cast_fp16 = mul(x = var_4765_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4767_cast_fp16")];
            int32 var_4769 = const()[name = string("op_4769"), val = int32(-2)];
            bool var_4770_interleave_0 = const()[name = string("op_4770_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_4770_cast_fp16 = concat(axis = var_4769, interleave = var_4770_interleave_0, values = (var_4767_cast_fp16, var_4765_cast_fp16_0))[name = string("op_4770_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4771_cast_fp16 = mul(x = var_4770_cast_fp16, y = var_459_cast_fp16)[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_81_cast_fp16 = add(x = var_4764_cast_fp16, y = var_4771_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4777_cast_fp16 = mul(x = var_4753_cast_fp16, y = var_452_cast_fp16)[name = string("op_4777_cast_fp16")];
            tensor<int32, [2]> var_4778_split_sizes_0 = const()[name = string("op_4778_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4778_axis_0 = const()[name = string("op_4778_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_4778_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_4778_cast_fp16_1 = split(axis = var_4778_axis_0, split_sizes = var_4778_split_sizes_0, x = var_4753_cast_fp16)[name = string("op_4778_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_4780_cast_fp16 = mul(x = var_4778_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4780_cast_fp16")];
            int32 var_4782 = const()[name = string("op_4782"), val = int32(-2)];
            bool var_4783_interleave_0 = const()[name = string("op_4783_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_4783_cast_fp16 = concat(axis = var_4782, interleave = var_4783_interleave_0, values = (var_4780_cast_fp16, var_4778_cast_fp16_0))[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4784_cast_fp16 = mul(x = var_4783_cast_fp16, y = var_459_cast_fp16)[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_135_cast_fp16 = add(x = var_4777_cast_fp16, y = var_4784_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_52)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_54_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_54 = read_state(input = key_cache)[name = string("coreml_update_state_54")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4760_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_53)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_55_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_55 = read_state(input = value_cache)[name = string("coreml_update_state_55")];
            tensor<int32, [4]> var_4854_begin_0 = const()[name = string("op_4854_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4854_end_0 = const()[name = string("op_4854_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4854_end_mask_0 = const()[name = string("op_4854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4854_cast_fp16 = slice_by_index(begin = var_4854_begin_0, end = var_4854_end_0, end_mask = var_4854_end_mask_0, x = coreml_update_state_54)[name = string("op_4854_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4857_axis_0 = const()[name = string("op_4857_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_1 = split(axis = var_4857_axis_0, split_sizes = tile_26, x = var_4854_cast_fp16)[name = string("op_4857_cast_fp16")];
            tensor<int32, [4]> var_4864_begin_0 = const()[name = string("op_4864_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4864_end_0 = const()[name = string("op_4864_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4864_end_mask_0 = const()[name = string("op_4864_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4864_cast_fp16 = slice_by_index(begin = var_4864_begin_0, end = var_4864_end_0, end_mask = var_4864_end_mask_0, x = coreml_update_state_55)[name = string("op_4864_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4867_axis_0 = const()[name = string("op_4867_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_1 = split(axis = var_4867_axis_0, split_sizes = tile_27, x = var_4864_cast_fp16)[name = string("op_4867_cast_fp16")];
            tensor<int32, [2]> var_4870_split_sizes_0 = const()[name = string("op_4870_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4870_axis_0 = const()[name = string("op_4870_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_4870_0, tensor<fp16, [1, 8, 128, 1]> var_4870_1 = split(axis = var_4870_axis_0, split_sizes = var_4870_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4870")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4857_cast_fp16_0, y = var_4870_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4873_to_fp16 = const()[name = string("op_4873_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4873_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4877 = const()[name = string("op_4877"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_215_cast_fp16 = softmax(axis = var_4877, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4883_transpose_x_1 = const()[name = string("op_4883_transpose_x_1"), val = bool(true)];
            bool var_4883_transpose_y_1 = const()[name = string("op_4883_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_4883_cast_fp16 = matmul(transpose_x = var_4883_transpose_x_1, transpose_y = var_4883_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4867_cast_fp16_0)[name = string("op_4883_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4857_cast_fp16_1, y = var_4870_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4885_to_fp16 = const()[name = string("op_4885_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4885_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4889 = const()[name = string("op_4889"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_cast_fp16 = softmax(axis = var_4889, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4867_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4897 = const()[name = string("op_4897"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_107_cast_fp16 = concat(axis = var_4897, interleave = attn_output_107_interleave_0, values = (var_4883_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4901_perm_0 = const()[name = string("op_4901_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_4901_cast_fp16 = transpose(perm = var_4901_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4901_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4934_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4934_cast_fp16")];
            int32 var_4932 = const()[name = string("op_4932"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_109_cast_fp16 = concat(axis = var_4932, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4934_cast_fp16))[name = string("doubled_109_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(881558912)))];
            fp16 var_4944_to_fp16 = const()[name = string("op_4944_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_4944_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_4955_split_sizes_0 = const()[name = string("op_4955_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4955_axis_0 = const()[name = string("op_4955_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4955_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4955_cast_fp16_1 = split(axis = var_4955_axis_0, split_sizes = var_4955_split_sizes_0, x = out_cast_fp16)[name = string("op_4955_cast_fp16")];
            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_13_mlp_gate_proj_weight_cast_fp16, x = var_4955_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4972_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4972_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881567168)))];
            tensor<int32, [2]> var_4978_strides_0 = const()[name = string("op_4978_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4978_pad_type_0 = const()[name = string("op_4978_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4978_pad_0 = const()[name = string("op_4978_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4978_dilations_0 = const()[name = string("op_4978_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4978_groups_0 = const()[name = string("op_4978_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4978_cast_fp16 = conv(dilations = var_4978_dilations_0, groups = var_4978_groups_0, pad = var_4978_pad_0, pad_type = var_4978_pad_type_0, strides = var_4978_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4955_cast_fp16_0)[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_cast_fp16 = mul(x = var_4972_cast_fp16, y = var_4978_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_13_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states = add(x = hidden_states_135_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_4987_cast_fp16")];
        } -> (hidden_states);
    func length_128<ios18>(tensor<fp32, [1, 2048, 1, 128]> inputs_embeds, state<tensor<fp16, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [128]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_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(64))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524992))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524416))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_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(525312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13120640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13108288))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_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(13126848))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651776))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651200))))[name = string("layers_2_self_attn_v_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(13652096))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26247424))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26235072))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_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(26253632))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26778560))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26777984))))[name = string("layers_3_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_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(26778880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30977408))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30973248))))[name = string("layers_3_self_attn_o_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(30979520))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43566656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43562496))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_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(43568768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093120))))[name = string("layers_4_self_attn_v_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(44094016))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48292544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48288384))))[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(48294656))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60889984))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60877632))))[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(60896192))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73491520))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73479168))))[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(73497728))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86084864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86080704))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_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(86086976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611328))))[name = string("layers_5_self_attn_v_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(86612224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90810752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90806592))))[name = string("layers_5_self_attn_o_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(90812864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103395840))))[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(103414400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997376))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_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(116003648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528000))))[name = string("layers_6_self_attn_v_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(116528896))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120727424))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120723264))))[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(120729536))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133324864))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133312512))))[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(133331072))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145926400))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145914048))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_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(145932608))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158519744))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158515584))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_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(158521856))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046784))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046208))))[name = string("layers_7_self_attn_v_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(159047104))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163245632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241472))))[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(163247744))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175843072))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175830720))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_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(175849280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176374208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176373632))))[name = string("layers_8_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_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(176374528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180573056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180568896))))[name = string("layers_8_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(180575168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193170496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193158144))))[name = string("layers_8_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(193176704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205772032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205759680))))[name = string("layers_8_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_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(205778240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218365376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218361216))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_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(218367488))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218892416))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218891840))))[name = string("layers_9_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_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(218892736))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223091264))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223087104))))[name = string("layers_9_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(223093376))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235688704))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235676352))))[name = string("layers_9_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(235694912))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248290240))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248277888))))[name = string("layers_9_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_9_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(248296448))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260883584))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260879424))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_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(260885696))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410624))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410048))))[name = string("layers_10_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_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(261410944))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265609472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265605312))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(265611584))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278206912))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278194560))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(278213120))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290808448))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290796096))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_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(290814656))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303401792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303397632))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(303403904))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307602432))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307598272))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_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(307604544))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308129472))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308128896))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_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(308129792))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654720))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654144))))[name = string("layers_11_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(308655040))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312853568))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312849408))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(312855680))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325451008))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325438656))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(325457216))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338052544))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040192))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_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(338058752))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350645888))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350641728))))[name = string("layers_11_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(350648000))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354846528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354842368))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_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(354848640))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355373568))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355372992))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_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(355373888))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898816))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898240))))[name = string("layers_12_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(355899136))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360097664))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360093504))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(360099776))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372695104))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372682752))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(372701312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385296640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385284288))))[name = string("layers_12_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_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(385302848))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397889984))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397885824))))[name = string("layers_12_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(397892096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402090624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402086464))))[name = string("layers_13_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_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(402092736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617088))))[name = string("layers_13_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_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(402617984))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142912))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142336))))[name = string("layers_13_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(403143232))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407341760))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407337600))))[name = string("layers_13_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_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(407343872))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419939200))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419926848))))[name = string("layers_13_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_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(419945408))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432532544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432528384))))[name = string("layers_13_mlp_down_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_424 = const()[name = string("op_424"), val = int32(0)];
            bool var_426_exclusive_0 = const()[name = string("op_426_exclusive_0"), val = bool(false)];
            bool var_426_reverse_0 = const()[name = string("op_426_reverse_0"), val = bool(false)];
            tensor<fp16, [128]> var_426_cast_fp16 = cumsum(axis = var_424, exclusive = var_426_exclusive_0, reverse = var_426_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_426_cast_fp16")];
            fp16 var_428_promoted_to_fp16 = const()[name = string("op_428_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [128]> position_offsets_cast_fp16 = sub(x = var_426_cast_fp16, y = var_428_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_431_axes_0 = const()[name = string("op_431_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 128]> var_431_cast_fp16 = expand_dims(axes = var_431_axes_0, x = position_offsets_cast_fp16)[name = string("op_431_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_431_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(432534656)))];
            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(440923328)))];
            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_450_perm_0 = const()[name = string("op_450_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_452_axes_0 = const()[name = string("op_452_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 128]> var_450_cast_fp16 = transpose(perm = var_450_perm_0, x = cos_1_cast_fp16)[name = string("transpose_314")];
            tensor<fp16, [1, 1, 128, 128]> var_452_cast_fp16 = expand_dims(axes = var_452_axes_0, x = var_450_cast_fp16)[name = string("op_452_cast_fp16")];
            tensor<int32, [3]> var_457_perm_0 = const()[name = string("op_457_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_459_axes_0 = const()[name = string("op_459_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 128]> var_457_cast_fp16 = transpose(perm = var_457_perm_0, x = sin_1_cast_fp16)[name = string("transpose_313")];
            tensor<fp16, [1, 1, 128, 128]> var_459_cast_fp16 = expand_dims(axes = var_459_axes_0, x = var_457_cast_fp16)[name = string("op_459_cast_fp16")];
            tensor<int32, [1]> var_478_axes_0 = const()[name = string("op_478_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 128, 1]> var_478 = expand_dims(axes = var_478_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_478")];
            tensor<int32, [1, 1, 2048]> var_471 = const()[name = string("op_471"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449312000)))];
            tensor<bool, [1, 128, 2048]> var_479 = greater(x = var_471, y = var_478)[name = string("op_479")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_486_axes_0 = const()[name = string("op_486_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 2048]> var_479_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_479)[name = string("cast_25")];
            tensor<fp16, [1, 1, 128, 2048]> var_486_cast_fp16 = expand_dims(axes = var_486_axes_0, x = var_479_to_fp16)[name = string("op_486_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_490_promoted_to_fp16 = const()[name = string("op_490_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 128]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_486_cast_fp16)[name = string("transpose_312")];
            tensor<bool, [1, 1, 2048, 128]> var_491_cast_fp16 = equal(x = mask_cast_fp16, y = var_490_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            fp16 var_492_to_fp16 = const()[name = string("op_492_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 128]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_492_to_fp16, cond = var_491_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_502_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_502_cast_fp16")];
            int32 var_500 = const()[name = string("op_500"), 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_500, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_502_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(449320256)))];
            fp16 var_512_to_fp16 = const()[name = string("op_512_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_512_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_523_split_sizes_0 = const()[name = string("op_523_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_523_axis_0 = const()[name = string("op_523_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_523_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_523_cast_fp16_1 = split(axis = var_523_axis_0, split_sizes = var_523_split_sizes_0, x = out_1_cast_fp16)[name = string("op_523_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449328512)))];
            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_to_fp16, x = var_523_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457717184)))];
            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_to_fp16, x = var_523_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(458765824)))];
            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_523_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_580_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_580_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_587_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_587_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_591_cast_fp16 = mul(x = x_1_cast_fp16, y = var_452_cast_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [2]> var_592_split_sizes_0 = const()[name = string("op_592_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_592_axis_0 = const()[name = string("op_592_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_592_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_592_cast_fp16_1 = split(axis = var_592_axis_0, split_sizes = var_592_split_sizes_0, x = x_1_cast_fp16)[name = string("op_592_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_594_cast_fp16 = mul(x = var_592_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_594_cast_fp16")];
            int32 var_596 = const()[name = string("op_596"), val = int32(-2)];
            bool var_597_interleave_0 = const()[name = string("op_597_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_597_cast_fp16 = concat(axis = var_596, interleave = var_597_interleave_0, values = (var_594_cast_fp16, var_592_cast_fp16_0))[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_598_cast_fp16 = mul(x = var_597_cast_fp16, y = var_459_cast_fp16)[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_3_cast_fp16 = add(x = var_591_cast_fp16, y = var_598_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_604_cast_fp16 = mul(x = var_580_cast_fp16, y = var_452_cast_fp16)[name = string("op_604_cast_fp16")];
            tensor<int32, [2]> var_605_split_sizes_0 = const()[name = string("op_605_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_605_axis_0 = const()[name = string("op_605_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_605_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_605_cast_fp16_1 = split(axis = var_605_axis_0, split_sizes = var_605_split_sizes_0, x = var_580_cast_fp16)[name = string("op_605_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_607_cast_fp16 = mul(x = var_605_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_607_cast_fp16")];
            int32 var_609 = const()[name = string("op_609"), val = int32(-2)];
            bool var_610_interleave_0 = const()[name = string("op_610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_610_cast_fp16 = concat(axis = var_609, interleave = var_610_interleave_0, values = (var_607_cast_fp16, var_605_cast_fp16_0))[name = string("op_610_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_611_cast_fp16 = mul(x = var_610_cast_fp16, y = var_459_cast_fp16)[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_5_cast_fp16 = add(x = var_604_cast_fp16, y = var_611_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 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_311")];
            tensor<fp16, [14, 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_168_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_168 = read_state(input = key_cache)[name = string("coreml_update_state_168")];
            tensor<fp16, [14, 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_587_cast_fp16)[name = string("transpose_310")];
            tensor<fp16, [14, 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_169_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_169 = read_state(input = value_cache)[name = string("coreml_update_state_169")];
            tensor<int32, [4]> var_681_begin_0 = const()[name = string("op_681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_681_end_0 = const()[name = string("op_681_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_681_end_mask_0 = const()[name = string("op_681_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_681_cast_fp16 = slice_by_index(begin = var_681_begin_0, end = var_681_end_0, end_mask = var_681_end_mask_0, x = coreml_update_state_168)[name = string("op_681_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_684_axis_0 = const()[name = string("op_684_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_1 = split(axis = var_684_axis_0, split_sizes = tile_0, x = var_681_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [4]> var_691_begin_0 = const()[name = string("op_691_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_691_end_0 = const()[name = string("op_691_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_691_end_mask_0 = const()[name = string("op_691_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = coreml_update_state_169)[name = string("op_691_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_694_axis_0 = const()[name = string("op_694_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_1 = split(axis = var_694_axis_0, split_sizes = tile_1, x = var_691_cast_fp16)[name = string("op_694_cast_fp16")];
            tensor<int32, [2]> var_697_split_sizes_0 = const()[name = string("op_697_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_697_axis_0 = const()[name = string("op_697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_697_0, tensor<fp16, [1, 8, 128, 128]> var_697_1 = split(axis = var_697_axis_0, split_sizes = var_697_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_697")];
            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_684_cast_fp16_0, y = var_697_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_700_to_fp16 = const()[name = string("op_700_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_700_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_704 = const()[name = string("op_704"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_7_cast_fp16 = softmax(axis = var_704, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_710_transpose_x_1 = const()[name = string("op_710_transpose_x_1"), val = bool(true)];
            bool var_710_transpose_y_1 = const()[name = string("op_710_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_710_cast_fp16 = matmul(transpose_x = var_710_transpose_x_1, transpose_y = var_710_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_694_cast_fp16_0)[name = string("op_710_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_684_cast_fp16_1, y = var_697_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_712_to_fp16 = const()[name = string("op_712_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_712_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_716 = const()[name = string("op_716"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_15_cast_fp16 = softmax(axis = var_716, 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_694_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_724 = const()[name = string("op_724"), 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_724, interleave = attn_output_3_interleave_0, values = (var_710_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_728_perm_0 = const()[name = string("op_728_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_728_cast_fp16 = transpose(perm = var_728_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_309")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_728_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(459814464)))];
            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_761_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_761_cast_fp16")];
            int32 var_759 = const()[name = string("op_759"), 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_759, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_761_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(468203136)))];
            fp16 var_771_to_fp16 = const()[name = string("op_771_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_771_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_782_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = out_3_cast_fp16)[name = string("op_782_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468211392)))];
            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_to_fp16, x = var_782_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_799_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_799_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493377280)))];
            tensor<int32, [2]> var_805_strides_0 = const()[name = string("op_805_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_805_pad_type_0 = const()[name = string("op_805_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_805_pad_0 = const()[name = string("op_805_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_805_dilations_0 = const()[name = string("op_805_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_805_groups_0 = const()[name = string("op_805_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_805_cast_fp16 = conv(dilations = var_805_dilations_0, groups = var_805_groups_0, pad = var_805_pad_0, pad_type = var_805_pad_type_0, strides = var_805_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_782_cast_fp16_0)[name = string("op_805_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_9_cast_fp16 = mul(x = var_799_cast_fp16, y = var_805_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518543168)))];
            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_to_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_823_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_823_cast_fp16")];
            int32 var_821 = const()[name = string("op_821"), 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_821, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_823_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(543709056)))];
            fp16 var_833_to_fp16 = const()[name = string("op_833_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_833_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_844_split_sizes_0 = const()[name = string("op_844_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_844_axis_0 = const()[name = string("op_844_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_844_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_844_cast_fp16_1 = split(axis = var_844_axis_0, split_sizes = var_844_split_sizes_0, x = out_5_cast_fp16)[name = string("op_844_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543717312)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552105984)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            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_cast_fp16, x = var_844_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_901_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_901_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_908_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_908_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_912_cast_fp16 = mul(x = x_11_cast_fp16, y = var_452_cast_fp16)[name = string("op_912_cast_fp16")];
            tensor<int32, [2]> var_913_split_sizes_0 = const()[name = string("op_913_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_913_axis_0 = const()[name = string("op_913_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_913_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_913_cast_fp16_1 = split(axis = var_913_axis_0, split_sizes = var_913_split_sizes_0, x = x_11_cast_fp16)[name = string("op_913_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_915_cast_fp16 = mul(x = var_913_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_915_cast_fp16")];
            int32 var_917 = const()[name = string("op_917"), val = int32(-2)];
            bool var_918_interleave_0 = const()[name = string("op_918_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_918_cast_fp16 = concat(axis = var_917, interleave = var_918_interleave_0, values = (var_915_cast_fp16, var_913_cast_fp16_0))[name = string("op_918_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_919_cast_fp16 = mul(x = var_918_cast_fp16, y = var_459_cast_fp16)[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_9_cast_fp16 = add(x = var_912_cast_fp16, y = var_919_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_925_cast_fp16 = mul(x = var_901_cast_fp16, y = var_452_cast_fp16)[name = string("op_925_cast_fp16")];
            tensor<int32, [2]> var_926_split_sizes_0 = const()[name = string("op_926_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_926_axis_0 = const()[name = string("op_926_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_926_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_926_cast_fp16_1 = split(axis = var_926_axis_0, split_sizes = var_926_split_sizes_0, x = var_901_cast_fp16)[name = string("op_926_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_928_cast_fp16 = mul(x = var_926_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_928_cast_fp16")];
            int32 var_930 = const()[name = string("op_930"), val = int32(-2)];
            bool var_931_interleave_0 = const()[name = string("op_931_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_931_cast_fp16 = concat(axis = var_930, interleave = var_931_interleave_0, values = (var_928_cast_fp16, var_926_cast_fp16_0))[name = string("op_931_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_932_cast_fp16 = mul(x = var_931_cast_fp16, y = var_459_cast_fp16)[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_15_cast_fp16 = add(x = var_925_cast_fp16, y = var_932_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_308")];
            tensor<fp16, [14, 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_168)[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_170_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_170 = read_state(input = key_cache)[name = string("coreml_update_state_170")];
            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_908_cast_fp16)[name = string("transpose_307")];
            tensor<fp16, [14, 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_169)[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_171_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_171 = read_state(input = value_cache)[name = string("coreml_update_state_171")];
            tensor<int32, [4]> var_1002_begin_0 = const()[name = string("op_1002_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1002_end_0 = const()[name = string("op_1002_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1002_end_mask_0 = const()[name = string("op_1002_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1002_cast_fp16 = slice_by_index(begin = var_1002_begin_0, end = var_1002_end_0, end_mask = var_1002_end_mask_0, x = coreml_update_state_170)[name = string("op_1002_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1005_axis_0 = const()[name = string("op_1005_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_1 = split(axis = var_1005_axis_0, split_sizes = tile_2, x = var_1002_cast_fp16)[name = string("op_1005_cast_fp16")];
            tensor<int32, [4]> var_1012_begin_0 = const()[name = string("op_1012_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1012_end_0 = const()[name = string("op_1012_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1012_end_mask_0 = const()[name = string("op_1012_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1012_cast_fp16 = slice_by_index(begin = var_1012_begin_0, end = var_1012_end_0, end_mask = var_1012_end_mask_0, x = coreml_update_state_171)[name = string("op_1012_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1015_axis_0 = const()[name = string("op_1015_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_1 = split(axis = var_1015_axis_0, split_sizes = tile_3, x = var_1012_cast_fp16)[name = string("op_1015_cast_fp16")];
            tensor<int32, [2]> var_1018_split_sizes_0 = const()[name = string("op_1018_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1018_axis_0 = const()[name = string("op_1018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1018_0, tensor<fp16, [1, 8, 128, 128]> var_1018_1 = split(axis = var_1018_axis_0, split_sizes = var_1018_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1018")];
            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_1005_cast_fp16_0, y = var_1018_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1021_to_fp16 = const()[name = string("op_1021_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_1021_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_1025 = const()[name = string("op_1025"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_23_cast_fp16 = softmax(axis = var_1025, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1031_transpose_x_1 = const()[name = string("op_1031_transpose_x_1"), val = bool(true)];
            bool var_1031_transpose_y_1 = const()[name = string("op_1031_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1031_cast_fp16 = matmul(transpose_x = var_1031_transpose_x_1, transpose_y = var_1031_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1015_cast_fp16_0)[name = string("op_1031_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_1005_cast_fp16_1, y = var_1018_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1033_to_fp16 = const()[name = string("op_1033_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_1033_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_1037 = const()[name = string("op_1037"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_31_cast_fp16 = softmax(axis = var_1037, 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_1015_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1045 = const()[name = string("op_1045"), 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_1045, interleave = attn_output_11_interleave_0, values = (var_1031_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1049_perm_0 = const()[name = string("op_1049_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_1049_cast_fp16 = transpose(perm = var_1049_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_306")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1049_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553154624)))];
            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_to_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_1082_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1082_cast_fp16")];
            int32 var_1080 = const()[name = string("op_1080"), 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_1080, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1082_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(561543296)))];
            fp16 var_1092_to_fp16 = const()[name = string("op_1092_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_1092_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1103_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1103_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561551552)))];
            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_to_fp16, x = var_1103_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_1120_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [2]> var_1126_strides_0 = const()[name = string("op_1126_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1126_pad_type_0 = const()[name = string("op_1126_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1126_pad_0 = const()[name = string("op_1126_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1126_dilations_0 = const()[name = string("op_1126_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1126_groups_0 = const()[name = string("op_1126_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_1126_cast_fp16 = conv(dilations = var_1126_dilations_0, groups = var_1126_groups_0, pad = var_1126_pad_0, pad_type = var_1126_pad_type_0, strides = var_1126_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1103_cast_fp16_0)[name = string("op_1126_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_19_cast_fp16 = mul(x = var_1120_cast_fp16, y = var_1126_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586717440)))];
            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_to_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_1144_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1144_cast_fp16")];
            int32 var_1142 = const()[name = string("op_1142"), 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_1142, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1144_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(611883328)))];
            fp16 var_1154_to_fp16 = const()[name = string("op_1154_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_1154_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1165_split_sizes_0 = const()[name = string("op_1165_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1165_axis_0 = const()[name = string("op_1165_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1165_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1165_cast_fp16_1 = split(axis = var_1165_axis_0, split_sizes = var_1165_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1165_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611891584)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620280256)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            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_cast_fp16, x = var_1165_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_1222_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1222_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_1229_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1229_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1233_cast_fp16 = mul(x = x_21_cast_fp16, y = var_452_cast_fp16)[name = string("op_1233_cast_fp16")];
            tensor<int32, [2]> var_1234_split_sizes_0 = const()[name = string("op_1234_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1234_axis_0 = const()[name = string("op_1234_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_1234_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_1234_cast_fp16_1 = split(axis = var_1234_axis_0, split_sizes = var_1234_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1234_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_1236_cast_fp16 = mul(x = var_1234_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1236_cast_fp16")];
            int32 var_1238 = const()[name = string("op_1238"), val = int32(-2)];
            bool var_1239_interleave_0 = const()[name = string("op_1239_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_1239_cast_fp16 = concat(axis = var_1238, interleave = var_1239_interleave_0, values = (var_1236_cast_fp16, var_1234_cast_fp16_0))[name = string("op_1239_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1240_cast_fp16 = mul(x = var_1239_cast_fp16, y = var_459_cast_fp16)[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_15_cast_fp16 = add(x = var_1233_cast_fp16, y = var_1240_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1246_cast_fp16 = mul(x = var_1222_cast_fp16, y = var_452_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<int32, [2]> var_1247_split_sizes_0 = const()[name = string("op_1247_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1247_axis_0 = const()[name = string("op_1247_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_1247_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_1247_cast_fp16_1 = split(axis = var_1247_axis_0, split_sizes = var_1247_split_sizes_0, x = var_1222_cast_fp16)[name = string("op_1247_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_1249_cast_fp16 = mul(x = var_1247_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1249_cast_fp16")];
            int32 var_1251 = const()[name = string("op_1251"), val = int32(-2)];
            bool var_1252_interleave_0 = const()[name = string("op_1252_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_1252_cast_fp16 = concat(axis = var_1251, interleave = var_1252_interleave_0, values = (var_1249_cast_fp16, var_1247_cast_fp16_0))[name = string("op_1252_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1253_cast_fp16 = mul(x = var_1252_cast_fp16, y = var_459_cast_fp16)[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_25_cast_fp16 = add(x = var_1246_cast_fp16, y = var_1253_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_305")];
            tensor<fp16, [14, 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_170)[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_172_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_172 = read_state(input = key_cache)[name = string("coreml_update_state_172")];
            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_1229_cast_fp16)[name = string("transpose_304")];
            tensor<fp16, [14, 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_171)[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_173_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_173 = read_state(input = value_cache)[name = string("coreml_update_state_173")];
            tensor<int32, [4]> var_1323_begin_0 = const()[name = string("op_1323_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1323_end_0 = const()[name = string("op_1323_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1323_end_mask_0 = const()[name = string("op_1323_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1323_cast_fp16 = slice_by_index(begin = var_1323_begin_0, end = var_1323_end_0, end_mask = var_1323_end_mask_0, x = coreml_update_state_172)[name = string("op_1323_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1326_axis_0 = const()[name = string("op_1326_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_1 = split(axis = var_1326_axis_0, split_sizes = tile_4, x = var_1323_cast_fp16)[name = string("op_1326_cast_fp16")];
            tensor<int32, [4]> var_1333_begin_0 = const()[name = string("op_1333_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1333_end_0 = const()[name = string("op_1333_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1333_end_mask_0 = const()[name = string("op_1333_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1333_cast_fp16 = slice_by_index(begin = var_1333_begin_0, end = var_1333_end_0, end_mask = var_1333_end_mask_0, x = coreml_update_state_173)[name = string("op_1333_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1336_axis_0 = const()[name = string("op_1336_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_1 = split(axis = var_1336_axis_0, split_sizes = tile_5, x = var_1333_cast_fp16)[name = string("op_1336_cast_fp16")];
            tensor<int32, [2]> var_1339_split_sizes_0 = const()[name = string("op_1339_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1339_axis_0 = const()[name = string("op_1339_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1339_0, tensor<fp16, [1, 8, 128, 128]> var_1339_1 = split(axis = var_1339_axis_0, split_sizes = var_1339_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1339")];
            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_1326_cast_fp16_0, y = var_1339_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1342_to_fp16 = const()[name = string("op_1342_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_1342_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_1346 = const()[name = string("op_1346"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_39_cast_fp16 = softmax(axis = var_1346, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1352_transpose_x_1 = const()[name = string("op_1352_transpose_x_1"), val = bool(true)];
            bool var_1352_transpose_y_1 = const()[name = string("op_1352_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1352_cast_fp16 = matmul(transpose_x = var_1352_transpose_x_1, transpose_y = var_1352_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1336_cast_fp16_0)[name = string("op_1352_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_1326_cast_fp16_1, y = var_1339_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1354_to_fp16 = const()[name = string("op_1354_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_1354_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_1358 = const()[name = string("op_1358"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_47_cast_fp16 = softmax(axis = var_1358, 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_1336_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1366 = const()[name = string("op_1366"), 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_1366, interleave = attn_output_19_interleave_0, values = (var_1352_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1370_perm_0 = const()[name = string("op_1370_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_1370_cast_fp16 = transpose(perm = var_1370_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_303")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1370_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621328896)))];
            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_to_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_1403_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1403_cast_fp16")];
            int32 var_1401 = const()[name = string("op_1401"), 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_1401, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1403_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(629717568)))];
            fp16 var_1413_to_fp16 = const()[name = string("op_1413_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_1413_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1424_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1424_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629725824)))];
            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_to_fp16, x = var_1424_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_1441_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1441_cast_fp16")];
            tensor<int32, [2]> var_1447_strides_0 = const()[name = string("op_1447_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1447_pad_type_0 = const()[name = string("op_1447_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1447_pad_0 = const()[name = string("op_1447_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1447_dilations_0 = const()[name = string("op_1447_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1447_groups_0 = const()[name = string("op_1447_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_1447_cast_fp16 = conv(dilations = var_1447_dilations_0, groups = var_1447_groups_0, pad = var_1447_pad_0, pad_type = var_1447_pad_type_0, strides = var_1447_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1424_cast_fp16_0)[name = string("op_1447_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_29_cast_fp16 = mul(x = var_1441_cast_fp16, y = var_1447_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654891712)))];
            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_to_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_1465_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1465_cast_fp16")];
            int32 var_1463 = const()[name = string("op_1463"), 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_1463, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1465_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(680057600)))];
            fp16 var_1475_to_fp16 = const()[name = string("op_1475_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_1475_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1486_split_sizes_0 = const()[name = string("op_1486_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1486_axis_0 = const()[name = string("op_1486_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1486_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1486_cast_fp16_1 = split(axis = var_1486_axis_0, split_sizes = var_1486_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1486_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680065856)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688454528)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            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_cast_fp16, x = var_1486_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_1543_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1543_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_1550_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1554_cast_fp16 = mul(x = x_31_cast_fp16, y = var_452_cast_fp16)[name = string("op_1554_cast_fp16")];
            tensor<int32, [2]> var_1555_split_sizes_0 = const()[name = string("op_1555_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1555_axis_0 = const()[name = string("op_1555_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_1555_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_1555_cast_fp16_1 = split(axis = var_1555_axis_0, split_sizes = var_1555_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1555_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_1557_cast_fp16 = mul(x = var_1555_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1557_cast_fp16")];
            int32 var_1559 = const()[name = string("op_1559"), val = int32(-2)];
            bool var_1560_interleave_0 = const()[name = string("op_1560_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_1560_cast_fp16 = concat(axis = var_1559, interleave = var_1560_interleave_0, values = (var_1557_cast_fp16, var_1555_cast_fp16_0))[name = string("op_1560_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1561_cast_fp16 = mul(x = var_1560_cast_fp16, y = var_459_cast_fp16)[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_21_cast_fp16 = add(x = var_1554_cast_fp16, y = var_1561_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1567_cast_fp16 = mul(x = var_1543_cast_fp16, y = var_452_cast_fp16)[name = string("op_1567_cast_fp16")];
            tensor<int32, [2]> var_1568_split_sizes_0 = const()[name = string("op_1568_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1568_axis_0 = const()[name = string("op_1568_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_1568_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_1568_cast_fp16_1 = split(axis = var_1568_axis_0, split_sizes = var_1568_split_sizes_0, x = var_1543_cast_fp16)[name = string("op_1568_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_1570_cast_fp16 = mul(x = var_1568_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1570_cast_fp16")];
            int32 var_1572 = const()[name = string("op_1572"), val = int32(-2)];
            bool var_1573_interleave_0 = const()[name = string("op_1573_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_1573_cast_fp16 = concat(axis = var_1572, interleave = var_1573_interleave_0, values = (var_1570_cast_fp16, var_1568_cast_fp16_0))[name = string("op_1573_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1574_cast_fp16 = mul(x = var_1573_cast_fp16, y = var_459_cast_fp16)[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_35_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1574_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_302")];
            tensor<fp16, [14, 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_172)[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_174_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_174 = read_state(input = key_cache)[name = string("coreml_update_state_174")];
            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_1550_cast_fp16)[name = string("transpose_301")];
            tensor<fp16, [14, 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_173)[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_175_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_175 = read_state(input = value_cache)[name = string("coreml_update_state_175")];
            tensor<int32, [4]> var_1644_begin_0 = const()[name = string("op_1644_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1644_end_0 = const()[name = string("op_1644_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1644_end_mask_0 = const()[name = string("op_1644_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1644_cast_fp16 = slice_by_index(begin = var_1644_begin_0, end = var_1644_end_0, end_mask = var_1644_end_mask_0, x = coreml_update_state_174)[name = string("op_1644_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1647_axis_0 = const()[name = string("op_1647_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_1 = split(axis = var_1647_axis_0, split_sizes = tile_6, x = var_1644_cast_fp16)[name = string("op_1647_cast_fp16")];
            tensor<int32, [4]> var_1654_begin_0 = const()[name = string("op_1654_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1654_end_0 = const()[name = string("op_1654_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1654_end_mask_0 = const()[name = string("op_1654_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1654_cast_fp16 = slice_by_index(begin = var_1654_begin_0, end = var_1654_end_0, end_mask = var_1654_end_mask_0, x = coreml_update_state_175)[name = string("op_1654_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1657_axis_0 = const()[name = string("op_1657_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_1 = split(axis = var_1657_axis_0, split_sizes = tile_7, x = var_1654_cast_fp16)[name = string("op_1657_cast_fp16")];
            tensor<int32, [2]> var_1660_split_sizes_0 = const()[name = string("op_1660_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1660_axis_0 = const()[name = string("op_1660_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1660_0, tensor<fp16, [1, 8, 128, 128]> var_1660_1 = split(axis = var_1660_axis_0, split_sizes = var_1660_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1660")];
            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_1647_cast_fp16_0, y = var_1660_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1663_to_fp16 = const()[name = string("op_1663_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_1663_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_1667 = const()[name = string("op_1667"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_55_cast_fp16 = softmax(axis = var_1667, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1673_transpose_x_1 = const()[name = string("op_1673_transpose_x_1"), val = bool(true)];
            bool var_1673_transpose_y_1 = const()[name = string("op_1673_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1673_cast_fp16 = matmul(transpose_x = var_1673_transpose_x_1, transpose_y = var_1673_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1657_cast_fp16_0)[name = string("op_1673_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_1647_cast_fp16_1, y = var_1660_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1675_to_fp16 = const()[name = string("op_1675_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_1675_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_1679 = const()[name = string("op_1679"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_63_cast_fp16 = softmax(axis = var_1679, 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_1657_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1687 = const()[name = string("op_1687"), 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_1687, interleave = attn_output_27_interleave_0, values = (var_1673_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1691_perm_0 = const()[name = string("op_1691_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_1691_cast_fp16 = transpose(perm = var_1691_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_300")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1691_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            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_cast_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_1724_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1724_cast_fp16")];
            int32 var_1722 = const()[name = string("op_1722"), 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_1722, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1724_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(689503168)))];
            fp16 var_1734_to_fp16 = const()[name = string("op_1734_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_1734_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1745_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1745_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689511424)))];
            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_to_fp16, x = var_1745_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_1762_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1762_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(714677312)))];
            tensor<int32, [2]> var_1768_strides_0 = const()[name = string("op_1768_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1768_pad_type_0 = const()[name = string("op_1768_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1768_pad_0 = const()[name = string("op_1768_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1768_dilations_0 = const()[name = string("op_1768_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1768_groups_0 = const()[name = string("op_1768_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_1768_cast_fp16 = conv(dilations = var_1768_dilations_0, groups = var_1768_groups_0, pad = var_1768_pad_0, pad_type = var_1768_pad_type_0, strides = var_1768_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1745_cast_fp16_0)[name = string("op_1768_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_39_cast_fp16 = mul(x = var_1762_cast_fp16, y = var_1768_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_1786_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1786_cast_fp16")];
            int32 var_1784 = const()[name = string("op_1784"), 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_1784, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1786_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(739843200)))];
            fp16 var_1796_to_fp16 = const()[name = string("op_1796_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_1796_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1807_split_sizes_0 = const()[name = string("op_1807_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1807_axis_0 = const()[name = string("op_1807_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_1807_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_1807_cast_fp16_1 = split(axis = var_1807_axis_0, split_sizes = var_1807_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739851456)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(748240128)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            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_cast_fp16, x = var_1807_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_1864_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1864_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_1871_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1871_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1875_cast_fp16 = mul(x = x_41_cast_fp16, y = var_452_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [2]> var_1876_split_sizes_0 = const()[name = string("op_1876_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1876_axis_0 = const()[name = string("op_1876_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_1876_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_1876_cast_fp16_1 = split(axis = var_1876_axis_0, split_sizes = var_1876_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1876_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_1878_cast_fp16 = mul(x = var_1876_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1878_cast_fp16")];
            int32 var_1880 = const()[name = string("op_1880"), val = int32(-2)];
            bool var_1881_interleave_0 = const()[name = string("op_1881_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_1881_cast_fp16 = concat(axis = var_1880, interleave = var_1881_interleave_0, values = (var_1878_cast_fp16, var_1876_cast_fp16_0))[name = string("op_1881_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1882_cast_fp16 = mul(x = var_1881_cast_fp16, y = var_459_cast_fp16)[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_27_cast_fp16 = add(x = var_1875_cast_fp16, y = var_1882_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1888_cast_fp16 = mul(x = var_1864_cast_fp16, y = var_452_cast_fp16)[name = string("op_1888_cast_fp16")];
            tensor<int32, [2]> var_1889_split_sizes_0 = const()[name = string("op_1889_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1889_axis_0 = const()[name = string("op_1889_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_1889_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_1889_cast_fp16_1 = split(axis = var_1889_axis_0, split_sizes = var_1889_split_sizes_0, x = var_1864_cast_fp16)[name = string("op_1889_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_1891_cast_fp16 = mul(x = var_1889_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1891_cast_fp16")];
            int32 var_1893 = const()[name = string("op_1893"), val = int32(-2)];
            bool var_1894_interleave_0 = const()[name = string("op_1894_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_1894_cast_fp16 = concat(axis = var_1893, interleave = var_1894_interleave_0, values = (var_1891_cast_fp16, var_1889_cast_fp16_0))[name = string("op_1894_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_1895_cast_fp16 = mul(x = var_1894_cast_fp16, y = var_459_cast_fp16)[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_45_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1895_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_299")];
            tensor<fp16, [14, 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_174)[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_176_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_176 = read_state(input = key_cache)[name = string("coreml_update_state_176")];
            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_1871_cast_fp16)[name = string("transpose_298")];
            tensor<fp16, [14, 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_175)[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_177_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_177 = read_state(input = value_cache)[name = string("coreml_update_state_177")];
            tensor<int32, [4]> var_1965_begin_0 = const()[name = string("op_1965_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1965_end_0 = const()[name = string("op_1965_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1965_end_mask_0 = const()[name = string("op_1965_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1965_cast_fp16 = slice_by_index(begin = var_1965_begin_0, end = var_1965_end_0, end_mask = var_1965_end_mask_0, x = coreml_update_state_176)[name = string("op_1965_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1968_axis_0 = const()[name = string("op_1968_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_1 = split(axis = var_1968_axis_0, split_sizes = tile_8, x = var_1965_cast_fp16)[name = string("op_1968_cast_fp16")];
            tensor<int32, [4]> var_1975_begin_0 = const()[name = string("op_1975_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1975_end_0 = const()[name = string("op_1975_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1975_end_mask_0 = const()[name = string("op_1975_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1975_cast_fp16 = slice_by_index(begin = var_1975_begin_0, end = var_1975_end_0, end_mask = var_1975_end_mask_0, x = coreml_update_state_177)[name = string("op_1975_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1978_axis_0 = const()[name = string("op_1978_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_1 = split(axis = var_1978_axis_0, split_sizes = tile_9, x = var_1975_cast_fp16)[name = string("op_1978_cast_fp16")];
            tensor<int32, [2]> var_1981_split_sizes_0 = const()[name = string("op_1981_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1981_axis_0 = const()[name = string("op_1981_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_1981_0, tensor<fp16, [1, 8, 128, 128]> var_1981_1 = split(axis = var_1981_axis_0, split_sizes = var_1981_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1981")];
            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_1968_cast_fp16_0, y = var_1981_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1984_to_fp16 = const()[name = string("op_1984_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_1984_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_1988 = const()[name = string("op_1988"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_71_cast_fp16 = softmax(axis = var_1988, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1994_transpose_x_1 = const()[name = string("op_1994_transpose_x_1"), val = bool(true)];
            bool var_1994_transpose_y_1 = const()[name = string("op_1994_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1994_cast_fp16 = matmul(transpose_x = var_1994_transpose_x_1, transpose_y = var_1994_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1978_cast_fp16_0)[name = string("op_1994_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_1968_cast_fp16_1, y = var_1981_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1996_to_fp16 = const()[name = string("op_1996_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_1996_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_2000 = const()[name = string("op_2000"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_79_cast_fp16 = softmax(axis = var_2000, 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_1978_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2008 = const()[name = string("op_2008"), 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_2008, interleave = attn_output_35_interleave_0, values = (var_1994_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2012_perm_0 = const()[name = string("op_2012_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_2012_cast_fp16 = transpose(perm = var_2012_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_297")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2012_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_2045_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2045_cast_fp16")];
            int32 var_2043 = const()[name = string("op_2043"), 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_2043, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2045_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(749288768)))];
            fp16 var_2055_to_fp16 = const()[name = string("op_2055_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_2055_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2066_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2066_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_2066_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_2083_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2083_cast_fp16")];
            tensor<int32, [2]> var_2089_strides_0 = const()[name = string("op_2089_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2089_pad_type_0 = const()[name = string("op_2089_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2089_pad_0 = const()[name = string("op_2089_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2089_dilations_0 = const()[name = string("op_2089_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2089_groups_0 = const()[name = string("op_2089_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_2089_cast_fp16 = conv(dilations = var_2089_dilations_0, groups = var_2089_groups_0, pad = var_2089_pad_0, pad_type = var_2089_pad_type_0, strides = var_2089_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2066_cast_fp16_0)[name = string("op_2089_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_49_cast_fp16 = mul(x = var_2083_cast_fp16, y = var_2089_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_2107_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2107_cast_fp16")];
            int32 var_2105 = const()[name = string("op_2105"), 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_2105, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2107_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(749297024)))];
            fp16 var_2117_to_fp16 = const()[name = string("op_2117_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_2117_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2128_split_sizes_0 = const()[name = string("op_2128_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2128_axis_0 = const()[name = string("op_2128_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2128_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2128_cast_fp16_1 = split(axis = var_2128_axis_0, split_sizes = var_2128_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749305280)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(757693952)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            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_cast_fp16, x = var_2128_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_2185_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2185_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_2192_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2196_cast_fp16 = mul(x = x_51_cast_fp16, y = var_452_cast_fp16)[name = string("op_2196_cast_fp16")];
            tensor<int32, [2]> var_2197_split_sizes_0 = const()[name = string("op_2197_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2197_axis_0 = const()[name = string("op_2197_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_2197_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_2197_cast_fp16_1 = split(axis = var_2197_axis_0, split_sizes = var_2197_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2197_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_2199_cast_fp16 = mul(x = var_2197_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2199_cast_fp16")];
            int32 var_2201 = const()[name = string("op_2201"), val = int32(-2)];
            bool var_2202_interleave_0 = const()[name = string("op_2202_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_2202_cast_fp16 = concat(axis = var_2201, interleave = var_2202_interleave_0, values = (var_2199_cast_fp16, var_2197_cast_fp16_0))[name = string("op_2202_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2203_cast_fp16 = mul(x = var_2202_cast_fp16, y = var_459_cast_fp16)[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_33_cast_fp16 = add(x = var_2196_cast_fp16, y = var_2203_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2209_cast_fp16 = mul(x = var_2185_cast_fp16, y = var_452_cast_fp16)[name = string("op_2209_cast_fp16")];
            tensor<int32, [2]> var_2210_split_sizes_0 = const()[name = string("op_2210_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2210_axis_0 = const()[name = string("op_2210_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_2210_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_2210_cast_fp16_1 = split(axis = var_2210_axis_0, split_sizes = var_2210_split_sizes_0, x = var_2185_cast_fp16)[name = string("op_2210_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_2212_cast_fp16 = mul(x = var_2210_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2212_cast_fp16")];
            int32 var_2214 = const()[name = string("op_2214"), val = int32(-2)];
            bool var_2215_interleave_0 = const()[name = string("op_2215_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_2215_cast_fp16 = concat(axis = var_2214, interleave = var_2215_interleave_0, values = (var_2212_cast_fp16, var_2210_cast_fp16_0))[name = string("op_2215_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2216_cast_fp16 = mul(x = var_2215_cast_fp16, y = var_459_cast_fp16)[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_55_cast_fp16 = add(x = var_2209_cast_fp16, y = var_2216_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_296")];
            tensor<fp16, [14, 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_176)[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_178_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_178 = read_state(input = key_cache)[name = string("coreml_update_state_178")];
            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_2192_cast_fp16)[name = string("transpose_295")];
            tensor<fp16, [14, 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_177)[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_179_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_179 = read_state(input = value_cache)[name = string("coreml_update_state_179")];
            tensor<int32, [4]> var_2286_begin_0 = const()[name = string("op_2286_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2286_end_0 = const()[name = string("op_2286_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2286_end_mask_0 = const()[name = string("op_2286_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2286_cast_fp16 = slice_by_index(begin = var_2286_begin_0, end = var_2286_end_0, end_mask = var_2286_end_mask_0, x = coreml_update_state_178)[name = string("op_2286_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2289_axis_0 = const()[name = string("op_2289_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_1 = split(axis = var_2289_axis_0, split_sizes = tile_10, x = var_2286_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> var_2296_begin_0 = const()[name = string("op_2296_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2296_end_0 = const()[name = string("op_2296_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2296_end_mask_0 = const()[name = string("op_2296_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2296_cast_fp16 = slice_by_index(begin = var_2296_begin_0, end = var_2296_end_0, end_mask = var_2296_end_mask_0, x = coreml_update_state_179)[name = string("op_2296_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2299_axis_0 = const()[name = string("op_2299_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_1 = split(axis = var_2299_axis_0, split_sizes = tile_11, x = var_2296_cast_fp16)[name = string("op_2299_cast_fp16")];
            tensor<int32, [2]> var_2302_split_sizes_0 = const()[name = string("op_2302_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2302_axis_0 = const()[name = string("op_2302_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_2302_0, tensor<fp16, [1, 8, 128, 128]> var_2302_1 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2302")];
            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_2289_cast_fp16_0, y = var_2302_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2305_to_fp16 = const()[name = string("op_2305_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_2305_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_2309 = const()[name = string("op_2309"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_87_cast_fp16 = softmax(axis = var_2309, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2315_transpose_x_1 = const()[name = string("op_2315_transpose_x_1"), val = bool(true)];
            bool var_2315_transpose_y_1 = const()[name = string("op_2315_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2315_cast_fp16 = matmul(transpose_x = var_2315_transpose_x_1, transpose_y = var_2315_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2299_cast_fp16_0)[name = string("op_2315_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_2289_cast_fp16_1, y = var_2302_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2317_to_fp16 = const()[name = string("op_2317_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_2317_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_2321 = const()[name = string("op_2321"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_95_cast_fp16 = softmax(axis = var_2321, 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_2299_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2329 = const()[name = string("op_2329"), 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_2329, interleave = attn_output_43_interleave_0, values = (var_2315_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2333_perm_0 = const()[name = string("op_2333_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_2333_cast_fp16 = transpose(perm = var_2333_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_294")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2333_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_2366_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2366_cast_fp16")];
            int32 var_2364 = const()[name = string("op_2364"), 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_2364, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2366_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(758742592)))];
            fp16 var_2376_to_fp16 = const()[name = string("op_2376_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_2376_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2387_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2387_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(758750848)))];
            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_to_fp16, x = var_2387_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_2404_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [2]> var_2410_strides_0 = const()[name = string("op_2410_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2410_pad_type_0 = const()[name = string("op_2410_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2410_pad_0 = const()[name = string("op_2410_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2410_dilations_0 = const()[name = string("op_2410_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2410_groups_0 = const()[name = string("op_2410_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_2410_cast_fp16 = conv(dilations = var_2410_dilations_0, groups = var_2410_groups_0, pad = var_2410_pad_0, pad_type = var_2410_pad_type_0, strides = var_2410_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2387_cast_fp16_0)[name = string("op_2410_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_59_cast_fp16 = mul(x = var_2404_cast_fp16, y = var_2410_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_2428_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2428_cast_fp16")];
            int32 var_2426 = const()[name = string("op_2426"), 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_2426, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2428_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(783916736)))];
            fp16 var_2438_to_fp16 = const()[name = string("op_2438_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_2438_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2449_split_sizes_0 = const()[name = string("op_2449_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2449_axis_0 = const()[name = string("op_2449_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2449_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2449_cast_fp16_1 = split(axis = var_2449_axis_0, split_sizes = var_2449_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2449_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(783924992)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792313664)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            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_cast_fp16, x = var_2449_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_2506_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2506_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_2513_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2517_cast_fp16 = mul(x = x_61_cast_fp16, y = var_452_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<int32, [2]> var_2518_split_sizes_0 = const()[name = string("op_2518_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2518_axis_0 = const()[name = string("op_2518_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_2518_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_2518_cast_fp16_1 = split(axis = var_2518_axis_0, split_sizes = var_2518_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2518_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_2520_cast_fp16 = mul(x = var_2518_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2520_cast_fp16")];
            int32 var_2522 = const()[name = string("op_2522"), val = int32(-2)];
            bool var_2523_interleave_0 = const()[name = string("op_2523_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_2523_cast_fp16 = concat(axis = var_2522, interleave = var_2523_interleave_0, values = (var_2520_cast_fp16, var_2518_cast_fp16_0))[name = string("op_2523_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2524_cast_fp16 = mul(x = var_2523_cast_fp16, y = var_459_cast_fp16)[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_39_cast_fp16 = add(x = var_2517_cast_fp16, y = var_2524_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2530_cast_fp16 = mul(x = var_2506_cast_fp16, y = var_452_cast_fp16)[name = string("op_2530_cast_fp16")];
            tensor<int32, [2]> var_2531_split_sizes_0 = const()[name = string("op_2531_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2531_axis_0 = const()[name = string("op_2531_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_2531_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_2531_cast_fp16_1 = split(axis = var_2531_axis_0, split_sizes = var_2531_split_sizes_0, x = var_2506_cast_fp16)[name = string("op_2531_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_2533_cast_fp16 = mul(x = var_2531_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2533_cast_fp16")];
            int32 var_2535 = const()[name = string("op_2535"), val = int32(-2)];
            bool var_2536_interleave_0 = const()[name = string("op_2536_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_2536_cast_fp16 = concat(axis = var_2535, interleave = var_2536_interleave_0, values = (var_2533_cast_fp16, var_2531_cast_fp16_0))[name = string("op_2536_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2537_cast_fp16 = mul(x = var_2536_cast_fp16, y = var_459_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_65_cast_fp16 = add(x = var_2530_cast_fp16, y = var_2537_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_293")];
            tensor<fp16, [14, 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_178)[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_180_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_180 = read_state(input = key_cache)[name = string("coreml_update_state_180")];
            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_2513_cast_fp16)[name = string("transpose_292")];
            tensor<fp16, [14, 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_179)[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_181_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_181 = read_state(input = value_cache)[name = string("coreml_update_state_181")];
            tensor<int32, [4]> var_2607_begin_0 = const()[name = string("op_2607_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2607_end_0 = const()[name = string("op_2607_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2607_end_mask_0 = const()[name = string("op_2607_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2607_cast_fp16 = slice_by_index(begin = var_2607_begin_0, end = var_2607_end_0, end_mask = var_2607_end_mask_0, x = coreml_update_state_180)[name = string("op_2607_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2610_axis_0 = const()[name = string("op_2610_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_1 = split(axis = var_2610_axis_0, split_sizes = tile_12, x = var_2607_cast_fp16)[name = string("op_2610_cast_fp16")];
            tensor<int32, [4]> var_2617_begin_0 = const()[name = string("op_2617_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2617_end_0 = const()[name = string("op_2617_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2617_end_mask_0 = const()[name = string("op_2617_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2617_cast_fp16 = slice_by_index(begin = var_2617_begin_0, end = var_2617_end_0, end_mask = var_2617_end_mask_0, x = coreml_update_state_181)[name = string("op_2617_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2620_axis_0 = const()[name = string("op_2620_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_1 = split(axis = var_2620_axis_0, split_sizes = tile_13, x = var_2617_cast_fp16)[name = string("op_2620_cast_fp16")];
            tensor<int32, [2]> var_2623_split_sizes_0 = const()[name = string("op_2623_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2623_axis_0 = const()[name = string("op_2623_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_2623_0, tensor<fp16, [1, 8, 128, 128]> var_2623_1 = split(axis = var_2623_axis_0, split_sizes = var_2623_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2623")];
            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_2610_cast_fp16_0, y = var_2623_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2626_to_fp16 = const()[name = string("op_2626_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_2626_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_2630 = const()[name = string("op_2630"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_103_cast_fp16 = softmax(axis = var_2630, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2636_transpose_x_1 = const()[name = string("op_2636_transpose_x_1"), val = bool(true)];
            bool var_2636_transpose_y_1 = const()[name = string("op_2636_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2636_cast_fp16 = matmul(transpose_x = var_2636_transpose_x_1, transpose_y = var_2636_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2620_cast_fp16_0)[name = string("op_2636_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_2610_cast_fp16_1, y = var_2623_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2638_to_fp16 = const()[name = string("op_2638_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_2638_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_2642 = const()[name = string("op_2642"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_111_cast_fp16 = softmax(axis = var_2642, 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_2620_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2650 = const()[name = string("op_2650"), 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_2650, interleave = attn_output_51_interleave_0, values = (var_2636_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2654_perm_0 = const()[name = string("op_2654_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_2654_cast_fp16 = transpose(perm = var_2654_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_291")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2654_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_2687_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2687_cast_fp16")];
            int32 var_2685 = const()[name = string("op_2685"), 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_2685, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2687_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(793362304)))];
            fp16 var_2697_to_fp16 = const()[name = string("op_2697_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_2697_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2708_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2708_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_2708_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_2725_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2725_cast_fp16")];
            tensor<int32, [2]> var_2731_strides_0 = const()[name = string("op_2731_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2731_pad_type_0 = const()[name = string("op_2731_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2731_pad_0 = const()[name = string("op_2731_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2731_dilations_0 = const()[name = string("op_2731_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2731_groups_0 = const()[name = string("op_2731_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_2731_cast_fp16 = conv(dilations = var_2731_dilations_0, groups = var_2731_groups_0, pad = var_2731_pad_0, pad_type = var_2731_pad_type_0, strides = var_2731_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2708_cast_fp16_0)[name = string("op_2731_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_69_cast_fp16 = mul(x = var_2725_cast_fp16, y = var_2731_cast_fp16)[name = string("x_69_cast_fp16")];
            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_cast_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_2749_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2749_cast_fp16")];
            int32 var_2747 = const()[name = string("op_2747"), 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_2747, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2749_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(793370560)))];
            fp16 var_2759_to_fp16 = const()[name = string("op_2759_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_2759_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2770_split_sizes_0 = const()[name = string("op_2770_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2770_axis_0 = const()[name = string("op_2770_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_2770_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_2770_cast_fp16_1 = split(axis = var_2770_axis_0, split_sizes = var_2770_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2770_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(793378816)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(801767488)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            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_cast_fp16, x = var_2770_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_2827_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2827_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_2834_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2834_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2838_cast_fp16 = mul(x = x_71_cast_fp16, y = var_452_cast_fp16)[name = string("op_2838_cast_fp16")];
            tensor<int32, [2]> var_2839_split_sizes_0 = const()[name = string("op_2839_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2839_axis_0 = const()[name = string("op_2839_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_2839_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_2839_cast_fp16_1 = split(axis = var_2839_axis_0, split_sizes = var_2839_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2839_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_2841_cast_fp16 = mul(x = var_2839_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2841_cast_fp16")];
            int32 var_2843 = const()[name = string("op_2843"), val = int32(-2)];
            bool var_2844_interleave_0 = const()[name = string("op_2844_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_2844_cast_fp16 = concat(axis = var_2843, interleave = var_2844_interleave_0, values = (var_2841_cast_fp16, var_2839_cast_fp16_0))[name = string("op_2844_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2845_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_459_cast_fp16)[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_45_cast_fp16 = add(x = var_2838_cast_fp16, y = var_2845_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2851_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_452_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [2]> var_2852_split_sizes_0 = const()[name = string("op_2852_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2852_axis_0 = const()[name = string("op_2852_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_2852_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_2852_cast_fp16_1 = split(axis = var_2852_axis_0, split_sizes = var_2852_split_sizes_0, x = var_2827_cast_fp16)[name = string("op_2852_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_2854_cast_fp16 = mul(x = var_2852_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2854_cast_fp16")];
            int32 var_2856 = const()[name = string("op_2856"), val = int32(-2)];
            bool var_2857_interleave_0 = const()[name = string("op_2857_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_2857_cast_fp16 = concat(axis = var_2856, interleave = var_2857_interleave_0, values = (var_2854_cast_fp16, var_2852_cast_fp16_0))[name = string("op_2857_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_2858_cast_fp16 = mul(x = var_2857_cast_fp16, y = var_459_cast_fp16)[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_75_cast_fp16 = add(x = var_2851_cast_fp16, y = var_2858_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_290")];
            tensor<fp16, [14, 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_180)[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_182_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_182 = read_state(input = key_cache)[name = string("coreml_update_state_182")];
            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_2834_cast_fp16)[name = string("transpose_289")];
            tensor<fp16, [14, 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_181)[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_183_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_183 = read_state(input = value_cache)[name = string("coreml_update_state_183")];
            tensor<int32, [4]> var_2928_begin_0 = const()[name = string("op_2928_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2928_end_0 = const()[name = string("op_2928_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2928_end_mask_0 = const()[name = string("op_2928_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2928_cast_fp16 = slice_by_index(begin = var_2928_begin_0, end = var_2928_end_0, end_mask = var_2928_end_mask_0, x = coreml_update_state_182)[name = string("op_2928_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2931_axis_0 = const()[name = string("op_2931_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_1 = split(axis = var_2931_axis_0, split_sizes = tile_14, x = var_2928_cast_fp16)[name = string("op_2931_cast_fp16")];
            tensor<int32, [4]> var_2938_begin_0 = const()[name = string("op_2938_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2938_end_0 = const()[name = string("op_2938_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2938_end_mask_0 = const()[name = string("op_2938_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2938_cast_fp16 = slice_by_index(begin = var_2938_begin_0, end = var_2938_end_0, end_mask = var_2938_end_mask_0, x = coreml_update_state_183)[name = string("op_2938_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2941_axis_0 = const()[name = string("op_2941_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_1 = split(axis = var_2941_axis_0, split_sizes = tile_15, x = var_2938_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<int32, [2]> var_2944_split_sizes_0 = const()[name = string("op_2944_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2944_axis_0 = const()[name = string("op_2944_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_2944_0, tensor<fp16, [1, 8, 128, 128]> var_2944_1 = split(axis = var_2944_axis_0, split_sizes = var_2944_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2944")];
            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_2931_cast_fp16_0, y = var_2944_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2947_to_fp16 = const()[name = string("op_2947_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_2947_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_2951 = const()[name = string("op_2951"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_119_cast_fp16 = softmax(axis = var_2951, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2957_transpose_x_1 = const()[name = string("op_2957_transpose_x_1"), val = bool(true)];
            bool var_2957_transpose_y_1 = const()[name = string("op_2957_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2957_cast_fp16 = matmul(transpose_x = var_2957_transpose_x_1, transpose_y = var_2957_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2941_cast_fp16_0)[name = string("op_2957_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_2931_cast_fp16_1, y = var_2944_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2959_to_fp16 = const()[name = string("op_2959_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_2959_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_2963 = const()[name = string("op_2963"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_127_cast_fp16 = softmax(axis = var_2963, 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_2941_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2971 = const()[name = string("op_2971"), 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_2971, interleave = attn_output_59_interleave_0, values = (var_2957_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2975_perm_0 = const()[name = string("op_2975_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_2975_cast_fp16 = transpose(perm = var_2975_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_288")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2975_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_3008_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3008_cast_fp16")];
            int32 var_3006 = const()[name = string("op_3006"), 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_3006, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3008_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(802816128)))];
            fp16 var_3018_to_fp16 = const()[name = string("op_3018_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_3018_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3029_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3029_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_3029_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_3046_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3046_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(802824384)))];
            tensor<int32, [2]> var_3052_strides_0 = const()[name = string("op_3052_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3052_pad_type_0 = const()[name = string("op_3052_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3052_pad_0 = const()[name = string("op_3052_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3052_dilations_0 = const()[name = string("op_3052_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3052_groups_0 = const()[name = string("op_3052_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_3052_cast_fp16 = conv(dilations = var_3052_dilations_0, groups = var_3052_groups_0, pad = var_3052_pad_0, pad_type = var_3052_pad_type_0, strides = var_3052_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_3029_cast_fp16_0)[name = string("op_3052_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_79_cast_fp16 = mul(x = var_3046_cast_fp16, y = var_3052_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(827990272)))];
            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_3070_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3070_cast_fp16")];
            int32 var_3068 = const()[name = string("op_3068"), 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_3068, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3070_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(853156160)))];
            fp16 var_3080_to_fp16 = const()[name = string("op_3080_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_3080_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3091_split_sizes_0 = const()[name = string("op_3091_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3091_axis_0 = const()[name = string("op_3091_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3091_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3091_cast_fp16_1 = split(axis = var_3091_axis_0, split_sizes = var_3091_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3091_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(853164416)))];
            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_3091_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(861553088)))];
            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_3091_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            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_cast_fp16, x = var_3091_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_3148_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3148_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_3155_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3155_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3159_cast_fp16 = mul(x = x_81_cast_fp16, y = var_452_cast_fp16)[name = string("op_3159_cast_fp16")];
            tensor<int32, [2]> var_3160_split_sizes_0 = const()[name = string("op_3160_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3160_axis_0 = const()[name = string("op_3160_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_3160_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_3160_cast_fp16_1 = split(axis = var_3160_axis_0, split_sizes = var_3160_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3160_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_3162_cast_fp16 = mul(x = var_3160_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3162_cast_fp16")];
            int32 var_3164 = const()[name = string("op_3164"), val = int32(-2)];
            bool var_3165_interleave_0 = const()[name = string("op_3165_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_3165_cast_fp16 = concat(axis = var_3164, interleave = var_3165_interleave_0, values = (var_3162_cast_fp16, var_3160_cast_fp16_0))[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3166_cast_fp16 = mul(x = var_3165_cast_fp16, y = var_459_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_51_cast_fp16 = add(x = var_3159_cast_fp16, y = var_3166_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3172_cast_fp16 = mul(x = var_3148_cast_fp16, y = var_452_cast_fp16)[name = string("op_3172_cast_fp16")];
            tensor<int32, [2]> var_3173_split_sizes_0 = const()[name = string("op_3173_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3173_axis_0 = const()[name = string("op_3173_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_3173_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_3173_cast_fp16_1 = split(axis = var_3173_axis_0, split_sizes = var_3173_split_sizes_0, x = var_3148_cast_fp16)[name = string("op_3173_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_3175_cast_fp16 = mul(x = var_3173_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3175_cast_fp16")];
            int32 var_3177 = const()[name = string("op_3177"), val = int32(-2)];
            bool var_3178_interleave_0 = const()[name = string("op_3178_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_3178_cast_fp16 = concat(axis = var_3177, interleave = var_3178_interleave_0, values = (var_3175_cast_fp16, var_3173_cast_fp16_0))[name = string("op_3178_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3179_cast_fp16 = mul(x = var_3178_cast_fp16, y = var_459_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_85_cast_fp16 = add(x = var_3172_cast_fp16, y = var_3179_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_287")];
            tensor<fp16, [14, 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_182)[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_184_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_184 = read_state(input = key_cache)[name = string("coreml_update_state_184")];
            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_3155_cast_fp16)[name = string("transpose_286")];
            tensor<fp16, [14, 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_183)[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_185_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_185 = read_state(input = value_cache)[name = string("coreml_update_state_185")];
            tensor<int32, [4]> var_3249_begin_0 = const()[name = string("op_3249_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3249_end_0 = const()[name = string("op_3249_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3249_end_mask_0 = const()[name = string("op_3249_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3249_cast_fp16 = slice_by_index(begin = var_3249_begin_0, end = var_3249_end_0, end_mask = var_3249_end_mask_0, x = coreml_update_state_184)[name = string("op_3249_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3252_axis_0 = const()[name = string("op_3252_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_1 = split(axis = var_3252_axis_0, split_sizes = tile_16, x = var_3249_cast_fp16)[name = string("op_3252_cast_fp16")];
            tensor<int32, [4]> var_3259_begin_0 = const()[name = string("op_3259_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3259_end_0 = const()[name = string("op_3259_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3259_end_mask_0 = const()[name = string("op_3259_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3259_cast_fp16 = slice_by_index(begin = var_3259_begin_0, end = var_3259_end_0, end_mask = var_3259_end_mask_0, x = coreml_update_state_185)[name = string("op_3259_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3262_axis_0 = const()[name = string("op_3262_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_1 = split(axis = var_3262_axis_0, split_sizes = tile_17, x = var_3259_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<int32, [2]> var_3265_split_sizes_0 = const()[name = string("op_3265_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3265_axis_0 = const()[name = string("op_3265_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_3265_0, tensor<fp16, [1, 8, 128, 128]> var_3265_1 = split(axis = var_3265_axis_0, split_sizes = var_3265_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3265")];
            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_3252_cast_fp16_0, y = var_3265_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3268_to_fp16 = const()[name = string("op_3268_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_3268_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_3272 = const()[name = string("op_3272"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_135_cast_fp16 = softmax(axis = var_3272, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3278_transpose_x_1 = const()[name = string("op_3278_transpose_x_1"), val = bool(true)];
            bool var_3278_transpose_y_1 = const()[name = string("op_3278_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_3278_cast_fp16 = matmul(transpose_x = var_3278_transpose_x_1, transpose_y = var_3278_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3262_cast_fp16_0)[name = string("op_3278_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_3252_cast_fp16_1, y = var_3265_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3280_to_fp16 = const()[name = string("op_3280_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_3280_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_3284 = const()[name = string("op_3284"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_143_cast_fp16 = softmax(axis = var_3284, x = attn_weights_141_cast_fp16)[name = string("attn_weights_143_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_143_cast_fp16, y = var_3262_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3292 = const()[name = string("op_3292"), 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_3292, interleave = attn_output_67_interleave_0, values = (var_3278_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3296_perm_0 = const()[name = string("op_3296_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_3296_cast_fp16 = transpose(perm = var_3296_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_285")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3296_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            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_cast_fp16, x = attn_output_71_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_3329_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3329_cast_fp16")];
            int32 var_3327 = const()[name = string("op_3327"), 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_3327, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3329_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(862601728)))];
            fp16 var_3339_to_fp16 = const()[name = string("op_3339_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_3339_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3350_split_sizes_0 = const()[name = string("op_3350_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3350_axis_0 = const()[name = string("op_3350_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3350_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3350_cast_fp16_1 = split(axis = var_3350_axis_0, split_sizes = var_3350_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3350_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_3367_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3367_cast_fp16")];
            tensor<int32, [2]> var_3373_strides_0 = const()[name = string("op_3373_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3373_pad_type_0 = const()[name = string("op_3373_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3373_pad_0 = const()[name = string("op_3373_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3373_dilations_0 = const()[name = string("op_3373_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3373_groups_0 = const()[name = string("op_3373_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_3373_cast_fp16 = conv(dilations = var_3373_dilations_0, groups = var_3373_groups_0, pad = var_3373_pad_0, pad_type = var_3373_pad_type_0, strides = var_3373_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("op_3373_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_89_cast_fp16 = mul(x = var_3367_cast_fp16, y = var_3373_cast_fp16)[name = string("x_89_cast_fp16")];
            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_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_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_3391_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3391_cast_fp16")];
            int32 var_3389 = const()[name = string("op_3389"), 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_3389, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3391_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_37_gamma_0_to_fp16 = const()[name = string("out_37_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862609984)))];
            fp16 var_3401_to_fp16 = const()[name = string("op_3401_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_3401_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3412_split_sizes_0 = const()[name = string("op_3412_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3412_axis_0 = const()[name = string("op_3412_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3412_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3412_cast_fp16_1 = split(axis = var_3412_axis_0, split_sizes = var_3412_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3412_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862618240)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(871006912)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3412_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_3469_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3469_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_3476_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3480_cast_fp16 = mul(x = x_91_cast_fp16, y = var_452_cast_fp16)[name = string("op_3480_cast_fp16")];
            tensor<int32, [2]> var_3481_split_sizes_0 = const()[name = string("op_3481_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3481_axis_0 = const()[name = string("op_3481_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_3481_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_3481_cast_fp16_1 = split(axis = var_3481_axis_0, split_sizes = var_3481_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3481_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_3483_cast_fp16 = mul(x = var_3481_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3483_cast_fp16")];
            int32 var_3485 = const()[name = string("op_3485"), val = int32(-2)];
            bool var_3486_interleave_0 = const()[name = string("op_3486_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_3486_cast_fp16 = concat(axis = var_3485, interleave = var_3486_interleave_0, values = (var_3483_cast_fp16, var_3481_cast_fp16_0))[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3487_cast_fp16 = mul(x = var_3486_cast_fp16, y = var_459_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_57_cast_fp16 = add(x = var_3480_cast_fp16, y = var_3487_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3493_cast_fp16 = mul(x = var_3469_cast_fp16, y = var_452_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [2]> var_3494_split_sizes_0 = const()[name = string("op_3494_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3494_axis_0 = const()[name = string("op_3494_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_3494_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_3494_cast_fp16_1 = split(axis = var_3494_axis_0, split_sizes = var_3494_split_sizes_0, x = var_3469_cast_fp16)[name = string("op_3494_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_3496_cast_fp16 = mul(x = var_3494_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3496_cast_fp16")];
            int32 var_3498 = const()[name = string("op_3498"), val = int32(-2)];
            bool var_3499_interleave_0 = const()[name = string("op_3499_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_3499_cast_fp16 = concat(axis = var_3498, interleave = var_3499_interleave_0, values = (var_3496_cast_fp16, var_3494_cast_fp16_0))[name = string("op_3499_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3500_cast_fp16 = mul(x = var_3499_cast_fp16, y = var_459_cast_fp16)[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_95_cast_fp16 = add(x = var_3493_cast_fp16, y = var_3500_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_284")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_184)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_186_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_186 = read_state(input = key_cache)[name = string("coreml_update_state_186")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3476_cast_fp16)[name = string("transpose_283")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_185)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_187_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_187 = read_state(input = value_cache)[name = string("coreml_update_state_187")];
            tensor<int32, [4]> var_3570_begin_0 = const()[name = string("op_3570_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3570_end_0 = const()[name = string("op_3570_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3570_end_mask_0 = const()[name = string("op_3570_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = coreml_update_state_186)[name = string("op_3570_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3573_axis_0 = const()[name = string("op_3573_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_1 = split(axis = var_3573_axis_0, split_sizes = tile_18, x = var_3570_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [4]> var_3580_begin_0 = const()[name = string("op_3580_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3580_end_0 = const()[name = string("op_3580_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3580_end_mask_0 = const()[name = string("op_3580_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3580_cast_fp16 = slice_by_index(begin = var_3580_begin_0, end = var_3580_end_0, end_mask = var_3580_end_mask_0, x = coreml_update_state_187)[name = string("op_3580_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3583_axis_0 = const()[name = string("op_3583_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_1 = split(axis = var_3583_axis_0, split_sizes = tile_19, x = var_3580_cast_fp16)[name = string("op_3583_cast_fp16")];
            tensor<int32, [2]> var_3586_split_sizes_0 = const()[name = string("op_3586_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3586_axis_0 = const()[name = string("op_3586_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_3586_0, tensor<fp16, [1, 8, 128, 128]> var_3586_1 = split(axis = var_3586_axis_0, split_sizes = var_3586_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3586")];
            bool attn_weights_145_transpose_x_0 = const()[name = string("attn_weights_145_transpose_x_0"), val = bool(false)];
            bool attn_weights_145_transpose_y_0 = const()[name = string("attn_weights_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_145_cast_fp16 = matmul(transpose_x = attn_weights_145_transpose_x_0, transpose_y = attn_weights_145_transpose_y_0, x = var_3573_cast_fp16_0, y = var_3586_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3589_to_fp16 = const()[name = string("op_3589_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_147_cast_fp16 = mul(x = attn_weights_145_cast_fp16, y = var_3589_to_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_149_cast_fp16 = add(x = attn_weights_147_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_149_cast_fp16")];
            int32 var_3593 = const()[name = string("op_3593"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_151_cast_fp16 = softmax(axis = var_3593, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3599_transpose_x_1 = const()[name = string("op_3599_transpose_x_1"), val = bool(true)];
            bool var_3599_transpose_y_1 = const()[name = string("op_3599_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_3599_cast_fp16 = matmul(transpose_x = var_3599_transpose_x_1, transpose_y = var_3599_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3583_cast_fp16_0)[name = string("op_3599_cast_fp16")];
            bool attn_weights_153_transpose_x_0 = const()[name = string("attn_weights_153_transpose_x_0"), val = bool(false)];
            bool attn_weights_153_transpose_y_0 = const()[name = string("attn_weights_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_153_cast_fp16 = matmul(transpose_x = attn_weights_153_transpose_x_0, transpose_y = attn_weights_153_transpose_y_0, x = var_3573_cast_fp16_1, y = var_3586_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3601_to_fp16 = const()[name = string("op_3601_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_155_cast_fp16 = mul(x = attn_weights_153_cast_fp16, y = var_3601_to_fp16)[name = string("attn_weights_155_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_157_cast_fp16 = add(x = attn_weights_155_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_157_cast_fp16")];
            int32 var_3605 = const()[name = string("op_3605"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_159_cast_fp16 = softmax(axis = var_3605, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3583_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3613 = const()[name = string("op_3613"), val = int32(1)];
            bool attn_output_75_interleave_0 = const()[name = string("attn_output_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_75_cast_fp16 = concat(axis = var_3613, interleave = attn_output_75_interleave_0, values = (var_3599_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3617_perm_0 = const()[name = string("op_3617_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_3617_cast_fp16 = transpose(perm = var_3617_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_282")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3617_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_3650_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3650_cast_fp16")];
            int32 var_3648 = const()[name = string("op_3648"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_77_cast_fp16 = concat(axis = var_3648, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3650_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872055552)))];
            fp16 var_3660_to_fp16 = const()[name = string("op_3660_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3660_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3671_split_sizes_0 = const()[name = string("op_3671_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3671_axis_0 = const()[name = string("op_3671_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3671_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3671_cast_fp16_1 = split(axis = var_3671_axis_0, split_sizes = var_3671_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3671_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_3688_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3688_cast_fp16")];
            tensor<int32, [2]> var_3694_strides_0 = const()[name = string("op_3694_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3694_pad_type_0 = const()[name = string("op_3694_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3694_pad_0 = const()[name = string("op_3694_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3694_dilations_0 = const()[name = string("op_3694_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3694_groups_0 = const()[name = string("op_3694_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_3694_cast_fp16 = conv(dilations = var_3694_dilations_0, groups = var_3694_groups_0, pad = var_3694_pad_0, pad_type = var_3694_pad_type_0, strides = var_3694_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("op_3694_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_99_cast_fp16 = mul(x = var_3688_cast_fp16, y = var_3694_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_3712_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3712_cast_fp16")];
            int32 var_3710 = const()[name = string("op_3710"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_81_cast_fp16 = concat(axis = var_3710, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3712_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872063808)))];
            fp16 var_3722_to_fp16 = const()[name = string("op_3722_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3722_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3733_split_sizes_0 = const()[name = string("op_3733_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3733_axis_0 = const()[name = string("op_3733_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3733_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3733_cast_fp16_1 = split(axis = var_3733_axis_0, split_sizes = var_3733_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872072064)))];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(880460736)))];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_cast_fp16, x = var_3733_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_3790_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3790_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_3797_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3797_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3801_cast_fp16 = mul(x = x_101_cast_fp16, y = var_452_cast_fp16)[name = string("op_3801_cast_fp16")];
            tensor<int32, [2]> var_3802_split_sizes_0 = const()[name = string("op_3802_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3802_axis_0 = const()[name = string("op_3802_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_3802_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_3802_cast_fp16_1 = split(axis = var_3802_axis_0, split_sizes = var_3802_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3802_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_3804_cast_fp16 = mul(x = var_3802_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3804_cast_fp16")];
            int32 var_3806 = const()[name = string("op_3806"), val = int32(-2)];
            bool var_3807_interleave_0 = const()[name = string("op_3807_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_3807_cast_fp16 = concat(axis = var_3806, interleave = var_3807_interleave_0, values = (var_3804_cast_fp16, var_3802_cast_fp16_0))[name = string("op_3807_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_3808_cast_fp16 = mul(x = var_3807_cast_fp16, y = var_459_cast_fp16)[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_63_cast_fp16 = add(x = var_3801_cast_fp16, y = var_3808_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3814_cast_fp16 = mul(x = var_3790_cast_fp16, y = var_452_cast_fp16)[name = string("op_3814_cast_fp16")];
            tensor<int32, [2]> var_3815_split_sizes_0 = const()[name = string("op_3815_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3815_axis_0 = const()[name = string("op_3815_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_3815_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_3815_cast_fp16_1 = split(axis = var_3815_axis_0, split_sizes = var_3815_split_sizes_0, x = var_3790_cast_fp16)[name = string("op_3815_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_3817_cast_fp16 = mul(x = var_3815_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3817_cast_fp16")];
            int32 var_3819 = const()[name = string("op_3819"), val = int32(-2)];
            bool var_3820_interleave_0 = const()[name = string("op_3820_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_3820_cast_fp16 = concat(axis = var_3819, interleave = var_3820_interleave_0, values = (var_3817_cast_fp16, var_3815_cast_fp16_0))[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_3821_cast_fp16 = mul(x = var_3820_cast_fp16, y = var_459_cast_fp16)[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_105_cast_fp16 = add(x = var_3814_cast_fp16, y = var_3821_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_281")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_186)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_188_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_188 = read_state(input = key_cache)[name = string("coreml_update_state_188")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3797_cast_fp16)[name = string("transpose_280")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_187)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_189_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_189 = read_state(input = value_cache)[name = string("coreml_update_state_189")];
            tensor<int32, [4]> var_3891_begin_0 = const()[name = string("op_3891_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3891_end_0 = const()[name = string("op_3891_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3891_end_mask_0 = const()[name = string("op_3891_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3891_cast_fp16 = slice_by_index(begin = var_3891_begin_0, end = var_3891_end_0, end_mask = var_3891_end_mask_0, x = coreml_update_state_188)[name = string("op_3891_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3894_axis_0 = const()[name = string("op_3894_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_1 = split(axis = var_3894_axis_0, split_sizes = tile_20, x = var_3891_cast_fp16)[name = string("op_3894_cast_fp16")];
            tensor<int32, [4]> var_3901_begin_0 = const()[name = string("op_3901_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3901_end_0 = const()[name = string("op_3901_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3901_end_mask_0 = const()[name = string("op_3901_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3901_cast_fp16 = slice_by_index(begin = var_3901_begin_0, end = var_3901_end_0, end_mask = var_3901_end_mask_0, x = coreml_update_state_189)[name = string("op_3901_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3904_axis_0 = const()[name = string("op_3904_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_1 = split(axis = var_3904_axis_0, split_sizes = tile_21, x = var_3901_cast_fp16)[name = string("op_3904_cast_fp16")];
            tensor<int32, [2]> var_3907_split_sizes_0 = const()[name = string("op_3907_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3907_axis_0 = const()[name = string("op_3907_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_3907_0, tensor<fp16, [1, 8, 128, 128]> var_3907_1 = split(axis = var_3907_axis_0, split_sizes = var_3907_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3907")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3894_cast_fp16_0, y = var_3907_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3910_to_fp16 = const()[name = string("op_3910_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3910_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3914 = const()[name = string("op_3914"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_167_cast_fp16 = softmax(axis = var_3914, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3920_transpose_x_1 = const()[name = string("op_3920_transpose_x_1"), val = bool(true)];
            bool var_3920_transpose_y_1 = const()[name = string("op_3920_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_3920_cast_fp16 = matmul(transpose_x = var_3920_transpose_x_1, transpose_y = var_3920_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3904_cast_fp16_0)[name = string("op_3920_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3894_cast_fp16_1, y = var_3907_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3922_to_fp16 = const()[name = string("op_3922_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3922_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3926 = const()[name = string("op_3926"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_175_cast_fp16 = softmax(axis = var_3926, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3904_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3934 = const()[name = string("op_3934"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_83_cast_fp16 = concat(axis = var_3934, interleave = attn_output_83_interleave_0, values = (var_3920_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3938_perm_0 = const()[name = string("op_3938_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_3938_cast_fp16 = transpose(perm = var_3938_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_279")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3938_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_3971_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3971_cast_fp16")];
            int32 var_3969 = const()[name = string("op_3969"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_85_cast_fp16 = concat(axis = var_3969, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3971_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881509376)))];
            fp16 var_3981_to_fp16 = const()[name = string("op_3981_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3981_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3992_split_sizes_0 = const()[name = string("op_3992_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3992_axis_0 = const()[name = string("op_3992_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_3992_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_3992_cast_fp16_1 = split(axis = var_3992_axis_0, split_sizes = var_3992_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3992_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_4009_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [2]> var_4015_strides_0 = const()[name = string("op_4015_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4015_pad_type_0 = const()[name = string("op_4015_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4015_pad_0 = const()[name = string("op_4015_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4015_dilations_0 = const()[name = string("op_4015_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4015_groups_0 = const()[name = string("op_4015_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_4015_cast_fp16 = conv(dilations = var_4015_dilations_0, groups = var_4015_groups_0, pad = var_4015_pad_0, pad_type = var_4015_pad_type_0, strides = var_4015_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("op_4015_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_109_cast_fp16 = mul(x = var_4009_cast_fp16, y = var_4015_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_4033_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4033_cast_fp16")];
            int32 var_4031 = const()[name = string("op_4031"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_89_cast_fp16 = concat(axis = var_4031, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4033_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881517632)))];
            fp16 var_4043_to_fp16 = const()[name = string("op_4043_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4043_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4054_split_sizes_0 = const()[name = string("op_4054_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4054_axis_0 = const()[name = string("op_4054_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_4054_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_4054_cast_fp16_1 = split(axis = var_4054_axis_0, split_sizes = var_4054_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_4111_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4111_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_4118_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4118_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_4122_cast_fp16 = mul(x = x_111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4122_cast_fp16")];
            tensor<int32, [2]> var_4123_split_sizes_0 = const()[name = string("op_4123_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4123_axis_0 = const()[name = string("op_4123_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_4123_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_4123_cast_fp16_1 = split(axis = var_4123_axis_0, split_sizes = var_4123_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4123_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_4125_cast_fp16 = mul(x = var_4123_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4125_cast_fp16")];
            int32 var_4127 = const()[name = string("op_4127"), val = int32(-2)];
            bool var_4128_interleave_0 = const()[name = string("op_4128_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_4128_cast_fp16 = concat(axis = var_4127, interleave = var_4128_interleave_0, values = (var_4125_cast_fp16, var_4123_cast_fp16_0))[name = string("op_4128_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_4129_cast_fp16 = mul(x = var_4128_cast_fp16, y = var_459_cast_fp16)[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_69_cast_fp16 = add(x = var_4122_cast_fp16, y = var_4129_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_4135_cast_fp16 = mul(x = var_4111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4135_cast_fp16")];
            tensor<int32, [2]> var_4136_split_sizes_0 = const()[name = string("op_4136_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4136_axis_0 = const()[name = string("op_4136_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_4136_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_4136_cast_fp16_1 = split(axis = var_4136_axis_0, split_sizes = var_4136_split_sizes_0, x = var_4111_cast_fp16)[name = string("op_4136_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_4138_cast_fp16 = mul(x = var_4136_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4138_cast_fp16")];
            int32 var_4140 = const()[name = string("op_4140"), val = int32(-2)];
            bool var_4141_interleave_0 = const()[name = string("op_4141_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_4141_cast_fp16 = concat(axis = var_4140, interleave = var_4141_interleave_0, values = (var_4138_cast_fp16, var_4136_cast_fp16_0))[name = string("op_4141_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_4142_cast_fp16 = mul(x = var_4141_cast_fp16, y = var_459_cast_fp16)[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_115_cast_fp16 = add(x = var_4135_cast_fp16, y = var_4142_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_278")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_188)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_190_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_190 = read_state(input = key_cache)[name = string("coreml_update_state_190")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4118_cast_fp16)[name = string("transpose_277")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_189)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_191_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_191 = read_state(input = value_cache)[name = string("coreml_update_state_191")];
            tensor<int32, [4]> var_4212_begin_0 = const()[name = string("op_4212_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4212_end_0 = const()[name = string("op_4212_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4212_end_mask_0 = const()[name = string("op_4212_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4212_cast_fp16 = slice_by_index(begin = var_4212_begin_0, end = var_4212_end_0, end_mask = var_4212_end_mask_0, x = coreml_update_state_190)[name = string("op_4212_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4215_axis_0 = const()[name = string("op_4215_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_1 = split(axis = var_4215_axis_0, split_sizes = tile_22, x = var_4212_cast_fp16)[name = string("op_4215_cast_fp16")];
            tensor<int32, [4]> var_4222_begin_0 = const()[name = string("op_4222_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4222_end_0 = const()[name = string("op_4222_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4222_end_mask_0 = const()[name = string("op_4222_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4222_cast_fp16 = slice_by_index(begin = var_4222_begin_0, end = var_4222_end_0, end_mask = var_4222_end_mask_0, x = coreml_update_state_191)[name = string("op_4222_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4225_axis_0 = const()[name = string("op_4225_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_1 = split(axis = var_4225_axis_0, split_sizes = tile_23, x = var_4222_cast_fp16)[name = string("op_4225_cast_fp16")];
            tensor<int32, [2]> var_4228_split_sizes_0 = const()[name = string("op_4228_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4228_axis_0 = const()[name = string("op_4228_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_4228_0, tensor<fp16, [1, 8, 128, 128]> var_4228_1 = split(axis = var_4228_axis_0, split_sizes = var_4228_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4228")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4215_cast_fp16_0, y = var_4228_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4231_to_fp16 = const()[name = string("op_4231_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4231_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4235 = const()[name = string("op_4235"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_183_cast_fp16 = softmax(axis = var_4235, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4241_transpose_x_1 = const()[name = string("op_4241_transpose_x_1"), val = bool(true)];
            bool var_4241_transpose_y_1 = const()[name = string("op_4241_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_4241_cast_fp16 = matmul(transpose_x = var_4241_transpose_x_1, transpose_y = var_4241_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4225_cast_fp16_0)[name = string("op_4241_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4215_cast_fp16_1, y = var_4228_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4243_to_fp16 = const()[name = string("op_4243_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4243_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4247 = const()[name = string("op_4247"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_191_cast_fp16 = softmax(axis = var_4247, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4225_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4255 = const()[name = string("op_4255"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_91_cast_fp16 = concat(axis = var_4255, interleave = attn_output_91_interleave_0, values = (var_4241_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4259_perm_0 = const()[name = string("op_4259_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_4259_cast_fp16 = transpose(perm = var_4259_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_276")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4259_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_4292_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4292_cast_fp16")];
            int32 var_4290 = const()[name = string("op_4290"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_93_cast_fp16 = concat(axis = var_4290, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4292_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881525888)))];
            fp16 var_4302_to_fp16 = const()[name = string("op_4302_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4302_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4313_split_sizes_0 = const()[name = string("op_4313_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4313_axis_0 = const()[name = string("op_4313_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_4313_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_4313_cast_fp16_1 = split(axis = var_4313_axis_0, split_sizes = var_4313_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4313_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_4330_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4330_cast_fp16")];
            tensor<int32, [2]> var_4336_strides_0 = const()[name = string("op_4336_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4336_pad_type_0 = const()[name = string("op_4336_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4336_pad_0 = const()[name = string("op_4336_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4336_dilations_0 = const()[name = string("op_4336_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4336_groups_0 = const()[name = string("op_4336_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_4336_cast_fp16 = conv(dilations = var_4336_dilations_0, groups = var_4336_groups_0, pad = var_4336_pad_0, pad_type = var_4336_pad_type_0, strides = var_4336_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("op_4336_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_119_cast_fp16 = mul(x = var_4330_cast_fp16, y = var_4336_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_cast_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_4354_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4354_cast_fp16")];
            int32 var_4352 = const()[name = string("op_4352"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_97_cast_fp16 = concat(axis = var_4352, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4354_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881534144)))];
            fp16 var_4364_to_fp16 = const()[name = string("op_4364_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4364_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4375_split_sizes_0 = const()[name = string("op_4375_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4375_axis_0 = const()[name = string("op_4375_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_4375_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_4375_cast_fp16_1 = split(axis = var_4375_axis_0, split_sizes = var_4375_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4375_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_4432_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4432_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_4439_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4439_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_4443_cast_fp16 = mul(x = x_121_cast_fp16, y = var_452_cast_fp16)[name = string("op_4443_cast_fp16")];
            tensor<int32, [2]> var_4444_split_sizes_0 = const()[name = string("op_4444_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4444_axis_0 = const()[name = string("op_4444_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_4444_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_4444_cast_fp16_1 = split(axis = var_4444_axis_0, split_sizes = var_4444_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4444_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_4446_cast_fp16 = mul(x = var_4444_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4446_cast_fp16")];
            int32 var_4448 = const()[name = string("op_4448"), val = int32(-2)];
            bool var_4449_interleave_0 = const()[name = string("op_4449_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_4449_cast_fp16 = concat(axis = var_4448, interleave = var_4449_interleave_0, values = (var_4446_cast_fp16, var_4444_cast_fp16_0))[name = string("op_4449_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_4450_cast_fp16 = mul(x = var_4449_cast_fp16, y = var_459_cast_fp16)[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_75_cast_fp16 = add(x = var_4443_cast_fp16, y = var_4450_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_4456_cast_fp16 = mul(x = var_4432_cast_fp16, y = var_452_cast_fp16)[name = string("op_4456_cast_fp16")];
            tensor<int32, [2]> var_4457_split_sizes_0 = const()[name = string("op_4457_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4457_axis_0 = const()[name = string("op_4457_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_4457_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_4457_cast_fp16_1 = split(axis = var_4457_axis_0, split_sizes = var_4457_split_sizes_0, x = var_4432_cast_fp16)[name = string("op_4457_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_4459_cast_fp16 = mul(x = var_4457_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4459_cast_fp16")];
            int32 var_4461 = const()[name = string("op_4461"), val = int32(-2)];
            bool var_4462_interleave_0 = const()[name = string("op_4462_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_4462_cast_fp16 = concat(axis = var_4461, interleave = var_4462_interleave_0, values = (var_4459_cast_fp16, var_4457_cast_fp16_0))[name = string("op_4462_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_4463_cast_fp16 = mul(x = var_4462_cast_fp16, y = var_459_cast_fp16)[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_125_cast_fp16 = add(x = var_4456_cast_fp16, y = var_4463_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_275")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_190)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_192_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_192 = read_state(input = key_cache)[name = string("coreml_update_state_192")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4439_cast_fp16)[name = string("transpose_274")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_191)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_193_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_193 = read_state(input = value_cache)[name = string("coreml_update_state_193")];
            tensor<int32, [4]> var_4533_begin_0 = const()[name = string("op_4533_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4533_end_0 = const()[name = string("op_4533_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4533_end_mask_0 = const()[name = string("op_4533_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4533_cast_fp16 = slice_by_index(begin = var_4533_begin_0, end = var_4533_end_0, end_mask = var_4533_end_mask_0, x = coreml_update_state_192)[name = string("op_4533_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4536_axis_0 = const()[name = string("op_4536_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_1 = split(axis = var_4536_axis_0, split_sizes = tile_24, x = var_4533_cast_fp16)[name = string("op_4536_cast_fp16")];
            tensor<int32, [4]> var_4543_begin_0 = const()[name = string("op_4543_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4543_end_0 = const()[name = string("op_4543_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4543_end_mask_0 = const()[name = string("op_4543_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4543_cast_fp16 = slice_by_index(begin = var_4543_begin_0, end = var_4543_end_0, end_mask = var_4543_end_mask_0, x = coreml_update_state_193)[name = string("op_4543_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4546_axis_0 = const()[name = string("op_4546_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_1 = split(axis = var_4546_axis_0, split_sizes = tile_25, x = var_4543_cast_fp16)[name = string("op_4546_cast_fp16")];
            tensor<int32, [2]> var_4549_split_sizes_0 = const()[name = string("op_4549_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4549_axis_0 = const()[name = string("op_4549_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_4549_0, tensor<fp16, [1, 8, 128, 128]> var_4549_1 = split(axis = var_4549_axis_0, split_sizes = var_4549_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4549")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4536_cast_fp16_0, y = var_4549_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4552_to_fp16 = const()[name = string("op_4552_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4552_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4556 = const()[name = string("op_4556"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_199_cast_fp16 = softmax(axis = var_4556, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4562_transpose_x_1 = const()[name = string("op_4562_transpose_x_1"), val = bool(true)];
            bool var_4562_transpose_y_1 = const()[name = string("op_4562_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_4562_cast_fp16 = matmul(transpose_x = var_4562_transpose_x_1, transpose_y = var_4562_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4546_cast_fp16_0)[name = string("op_4562_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4536_cast_fp16_1, y = var_4549_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4564_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4568 = const()[name = string("op_4568"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_207_cast_fp16 = softmax(axis = var_4568, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4546_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4576 = const()[name = string("op_4576"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_99_cast_fp16 = concat(axis = var_4576, interleave = attn_output_99_interleave_0, values = (var_4562_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4580_perm_0 = const()[name = string("op_4580_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_4580_cast_fp16 = transpose(perm = var_4580_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_273")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4580_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_4613_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4613_cast_fp16")];
            int32 var_4611 = const()[name = string("op_4611"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_101_cast_fp16 = concat(axis = var_4611, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4613_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881542400)))];
            fp16 var_4623_to_fp16 = const()[name = string("op_4623_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4623_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4634_split_sizes_0 = const()[name = string("op_4634_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4634_axis_0 = const()[name = string("op_4634_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_4634_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_4634_cast_fp16_1 = split(axis = var_4634_axis_0, split_sizes = var_4634_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4634_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_4651_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4651_cast_fp16")];
            tensor<int32, [2]> var_4657_strides_0 = const()[name = string("op_4657_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4657_pad_type_0 = const()[name = string("op_4657_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4657_pad_0 = const()[name = string("op_4657_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4657_dilations_0 = const()[name = string("op_4657_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4657_groups_0 = const()[name = string("op_4657_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_4657_cast_fp16 = conv(dilations = var_4657_dilations_0, groups = var_4657_groups_0, pad = var_4657_pad_0, pad_type = var_4657_pad_type_0, strides = var_4657_strides_0, weight = layers_12_mlp_up_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("op_4657_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_129_cast_fp16 = mul(x = var_4651_cast_fp16, y = var_4657_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_cast_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_4675_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4675_cast_fp16")];
            int32 var_4673 = const()[name = string("op_4673"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_105_cast_fp16 = concat(axis = var_4673, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4675_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881550656)))];
            fp16 var_4685_to_fp16 = const()[name = string("op_4685_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4685_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4696_split_sizes_0 = const()[name = string("op_4696_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4696_axis_0 = const()[name = string("op_4696_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_4696_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_4696_cast_fp16_1 = split(axis = var_4696_axis_0, split_sizes = var_4696_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4696_cast_fp16")];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 128]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 128]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_4753_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4753_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 128]> var_4760_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4760_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_4764_cast_fp16 = mul(x = x_131_cast_fp16, y = var_452_cast_fp16)[name = string("op_4764_cast_fp16")];
            tensor<int32, [2]> var_4765_split_sizes_0 = const()[name = string("op_4765_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4765_axis_0 = const()[name = string("op_4765_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 128]> var_4765_cast_fp16_0, tensor<fp16, [1, 16, 64, 128]> var_4765_cast_fp16_1 = split(axis = var_4765_axis_0, split_sizes = var_4765_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4765_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 128]> var_4767_cast_fp16 = mul(x = var_4765_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4767_cast_fp16")];
            int32 var_4769 = const()[name = string("op_4769"), val = int32(-2)];
            bool var_4770_interleave_0 = const()[name = string("op_4770_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> var_4770_cast_fp16 = concat(axis = var_4769, interleave = var_4770_interleave_0, values = (var_4767_cast_fp16, var_4765_cast_fp16_0))[name = string("op_4770_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_4771_cast_fp16 = mul(x = var_4770_cast_fp16, y = var_459_cast_fp16)[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> query_states_81_cast_fp16 = add(x = var_4764_cast_fp16, y = var_4771_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_4777_cast_fp16 = mul(x = var_4753_cast_fp16, y = var_452_cast_fp16)[name = string("op_4777_cast_fp16")];
            tensor<int32, [2]> var_4778_split_sizes_0 = const()[name = string("op_4778_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4778_axis_0 = const()[name = string("op_4778_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 128]> var_4778_cast_fp16_0, tensor<fp16, [1, 2, 64, 128]> var_4778_cast_fp16_1 = split(axis = var_4778_axis_0, split_sizes = var_4778_split_sizes_0, x = var_4753_cast_fp16)[name = string("op_4778_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_4780_cast_fp16 = mul(x = var_4778_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4780_cast_fp16")];
            int32 var_4782 = const()[name = string("op_4782"), val = int32(-2)];
            bool var_4783_interleave_0 = const()[name = string("op_4783_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 128]> var_4783_cast_fp16 = concat(axis = var_4782, interleave = var_4783_interleave_0, values = (var_4780_cast_fp16, var_4778_cast_fp16_0))[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> var_4784_cast_fp16 = mul(x = var_4783_cast_fp16, y = var_459_cast_fp16)[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 128]> key_states_135_cast_fp16 = add(x = var_4777_cast_fp16, y = var_4784_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_272")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_192)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_194_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_194 = read_state(input = key_cache)[name = string("coreml_update_state_194")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 128, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4760_cast_fp16)[name = string("transpose_271")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_193)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_195_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_195 = read_state(input = value_cache)[name = string("coreml_update_state_195")];
            tensor<int32, [4]> var_4854_begin_0 = const()[name = string("op_4854_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4854_end_0 = const()[name = string("op_4854_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4854_end_mask_0 = const()[name = string("op_4854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4854_cast_fp16 = slice_by_index(begin = var_4854_begin_0, end = var_4854_end_0, end_mask = var_4854_end_mask_0, x = coreml_update_state_194)[name = string("op_4854_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4857_axis_0 = const()[name = string("op_4857_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_1 = split(axis = var_4857_axis_0, split_sizes = tile_26, x = var_4854_cast_fp16)[name = string("op_4857_cast_fp16")];
            tensor<int32, [4]> var_4864_begin_0 = const()[name = string("op_4864_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4864_end_0 = const()[name = string("op_4864_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4864_end_mask_0 = const()[name = string("op_4864_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4864_cast_fp16 = slice_by_index(begin = var_4864_begin_0, end = var_4864_end_0, end_mask = var_4864_end_mask_0, x = coreml_update_state_195)[name = string("op_4864_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4867_axis_0 = const()[name = string("op_4867_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_1 = split(axis = var_4867_axis_0, split_sizes = tile_27, x = var_4864_cast_fp16)[name = string("op_4867_cast_fp16")];
            tensor<int32, [2]> var_4870_split_sizes_0 = const()[name = string("op_4870_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4870_axis_0 = const()[name = string("op_4870_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 128]> var_4870_0, tensor<fp16, [1, 8, 128, 128]> var_4870_1 = split(axis = var_4870_axis_0, split_sizes = var_4870_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4870")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4857_cast_fp16_0, y = var_4870_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4873_to_fp16 = const()[name = string("op_4873_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4873_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4877 = const()[name = string("op_4877"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_215_cast_fp16 = softmax(axis = var_4877, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4883_transpose_x_1 = const()[name = string("op_4883_transpose_x_1"), val = bool(true)];
            bool var_4883_transpose_y_1 = const()[name = string("op_4883_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_4883_cast_fp16 = matmul(transpose_x = var_4883_transpose_x_1, transpose_y = var_4883_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4867_cast_fp16_0)[name = string("op_4883_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4857_cast_fp16_1, y = var_4870_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4885_to_fp16 = const()[name = string("op_4885_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4885_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4889 = const()[name = string("op_4889"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 128]> attn_weights_cast_fp16 = softmax(axis = var_4889, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4867_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4897 = const()[name = string("op_4897"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> attn_output_107_cast_fp16 = concat(axis = var_4897, interleave = attn_output_107_interleave_0, values = (var_4883_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4901_perm_0 = const()[name = string("op_4901_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 128]> var_4901_cast_fp16 = transpose(perm = var_4901_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_270")];
            tensor<fp16, [1, 2048, 1, 128]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4901_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 128]> var_4934_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4934_cast_fp16")];
            int32 var_4932 = const()[name = string("op_4932"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 128]> doubled_109_cast_fp16 = concat(axis = var_4932, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4934_cast_fp16))[name = string("doubled_109_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(881558912)))];
            fp16 var_4944_to_fp16 = const()[name = string("op_4944_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 128]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_4944_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_4955_split_sizes_0 = const()[name = string("op_4955_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4955_axis_0 = const()[name = string("op_4955_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> var_4955_cast_fp16_0, tensor<fp16, [1, 2048, 1, 128]> var_4955_cast_fp16_1 = split(axis = var_4955_axis_0, split_sizes = var_4955_split_sizes_0, x = out_cast_fp16)[name = string("op_4955_cast_fp16")];
            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_13_mlp_gate_proj_weight_cast_fp16, x = var_4955_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> var_4972_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4972_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881567168)))];
            tensor<int32, [2]> var_4978_strides_0 = const()[name = string("op_4978_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4978_pad_type_0 = const()[name = string("op_4978_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4978_pad_0 = const()[name = string("op_4978_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4978_dilations_0 = const()[name = string("op_4978_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4978_groups_0 = const()[name = string("op_4978_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 128]> var_4978_cast_fp16 = conv(dilations = var_4978_dilations_0, groups = var_4978_groups_0, pad = var_4978_pad_0, pad_type = var_4978_pad_type_0, strides = var_4978_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4955_cast_fp16_0)[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 128]> x_cast_fp16 = mul(x = var_4972_cast_fp16, y = var_4978_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_13_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 128]> hidden_states = add(x = hidden_states_135_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_4987_cast_fp16")];
        } -> (hidden_states);
    func length_16<ios18>(tensor<fp32, [1, 2048, 1, 16]> inputs_embeds, state<tensor<fp16, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [16]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_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(64))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524992))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524416))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_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(525312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13120640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13108288))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_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(13126848))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651776))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651200))))[name = string("layers_2_self_attn_v_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(13652096))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26247424))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26235072))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_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(26253632))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26778560))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26777984))))[name = string("layers_3_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_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(26778880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30977408))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30973248))))[name = string("layers_3_self_attn_o_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(30979520))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43566656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43562496))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_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(43568768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093120))))[name = string("layers_4_self_attn_v_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(44094016))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48292544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48288384))))[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(48294656))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60889984))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60877632))))[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(60896192))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73491520))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73479168))))[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(73497728))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86084864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86080704))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_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(86086976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611328))))[name = string("layers_5_self_attn_v_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(86612224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90810752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90806592))))[name = string("layers_5_self_attn_o_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(90812864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103395840))))[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(103414400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997376))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_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(116003648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528000))))[name = string("layers_6_self_attn_v_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(116528896))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120727424))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120723264))))[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(120729536))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133324864))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133312512))))[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(133331072))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145926400))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145914048))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_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(145932608))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158519744))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158515584))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_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(158521856))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046784))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046208))))[name = string("layers_7_self_attn_v_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(159047104))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163245632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241472))))[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(163247744))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175843072))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175830720))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_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(175849280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176374208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176373632))))[name = string("layers_8_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_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(176374528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180573056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180568896))))[name = string("layers_8_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(180575168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193170496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193158144))))[name = string("layers_8_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(193176704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205772032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205759680))))[name = string("layers_8_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_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(205778240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218365376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218361216))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_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(218367488))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218892416))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218891840))))[name = string("layers_9_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_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(218892736))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223091264))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223087104))))[name = string("layers_9_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(223093376))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235688704))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235676352))))[name = string("layers_9_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(235694912))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248290240))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248277888))))[name = string("layers_9_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_9_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(248296448))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260883584))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260879424))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_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(260885696))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410624))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410048))))[name = string("layers_10_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_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(261410944))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265609472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265605312))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(265611584))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278206912))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278194560))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(278213120))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290808448))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290796096))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_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(290814656))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303401792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303397632))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(303403904))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307602432))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307598272))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_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(307604544))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308129472))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308128896))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_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(308129792))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654720))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654144))))[name = string("layers_11_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(308655040))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312853568))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312849408))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(312855680))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325451008))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325438656))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(325457216))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338052544))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040192))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_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(338058752))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350645888))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350641728))))[name = string("layers_11_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(350648000))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354846528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354842368))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_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(354848640))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355373568))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355372992))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_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(355373888))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898816))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898240))))[name = string("layers_12_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(355899136))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360097664))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360093504))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(360099776))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372695104))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372682752))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(372701312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385296640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385284288))))[name = string("layers_12_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_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(385302848))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397889984))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397885824))))[name = string("layers_12_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(397892096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402090624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402086464))))[name = string("layers_13_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_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(402092736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617088))))[name = string("layers_13_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_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(402617984))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142912))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142336))))[name = string("layers_13_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(403143232))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407341760))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407337600))))[name = string("layers_13_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_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(407343872))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419939200))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419926848))))[name = string("layers_13_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_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(419945408))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432532544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432528384))))[name = string("layers_13_mlp_down_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_424 = const()[name = string("op_424"), val = int32(0)];
            bool var_426_exclusive_0 = const()[name = string("op_426_exclusive_0"), val = bool(false)];
            bool var_426_reverse_0 = const()[name = string("op_426_reverse_0"), val = bool(false)];
            tensor<fp16, [16]> var_426_cast_fp16 = cumsum(axis = var_424, exclusive = var_426_exclusive_0, reverse = var_426_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_426_cast_fp16")];
            fp16 var_428_promoted_to_fp16 = const()[name = string("op_428_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [16]> position_offsets_cast_fp16 = sub(x = var_426_cast_fp16, y = var_428_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_431_axes_0 = const()[name = string("op_431_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 16]> var_431_cast_fp16 = expand_dims(axes = var_431_axes_0, x = position_offsets_cast_fp16)[name = string("op_431_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_431_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(432534656)))];
            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(440923328)))];
            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_450_perm_0 = const()[name = string("op_450_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_452_axes_0 = const()[name = string("op_452_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 16]> var_450_cast_fp16 = transpose(perm = var_450_perm_0, x = cos_1_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 1, 128, 16]> var_452_cast_fp16 = expand_dims(axes = var_452_axes_0, x = var_450_cast_fp16)[name = string("op_452_cast_fp16")];
            tensor<int32, [3]> var_457_perm_0 = const()[name = string("op_457_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_459_axes_0 = const()[name = string("op_459_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 16]> var_457_cast_fp16 = transpose(perm = var_457_perm_0, x = sin_1_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [1, 1, 128, 16]> var_459_cast_fp16 = expand_dims(axes = var_459_axes_0, x = var_457_cast_fp16)[name = string("op_459_cast_fp16")];
            tensor<int32, [1]> var_478_axes_0 = const()[name = string("op_478_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 16, 1]> var_478 = expand_dims(axes = var_478_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_478")];
            tensor<int32, [1, 1, 2048]> var_471 = const()[name = string("op_471"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449312000)))];
            tensor<bool, [1, 16, 2048]> var_479 = greater(x = var_471, y = var_478)[name = string("op_479")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_486_axes_0 = const()[name = string("op_486_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 16, 2048]> var_479_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_479)[name = string("cast_13")];
            tensor<fp16, [1, 1, 16, 2048]> var_486_cast_fp16 = expand_dims(axes = var_486_axes_0, x = var_479_to_fp16)[name = string("op_486_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_490_promoted_to_fp16 = const()[name = string("op_490_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 16]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_486_cast_fp16)[name = string("transpose_177")];
            tensor<bool, [1, 1, 2048, 16]> var_491_cast_fp16 = equal(x = mask_cast_fp16, y = var_490_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            fp16 var_492_to_fp16 = const()[name = string("op_492_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 16]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_492_to_fp16, cond = var_491_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_502_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_502_cast_fp16")];
            int32 var_500 = const()[name = string("op_500"), 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_500, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_502_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(449320256)))];
            fp16 var_512_to_fp16 = const()[name = string("op_512_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_512_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_523_split_sizes_0 = const()[name = string("op_523_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_523_axis_0 = const()[name = string("op_523_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_523_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_523_cast_fp16_1 = split(axis = var_523_axis_0, split_sizes = var_523_split_sizes_0, x = out_1_cast_fp16)[name = string("op_523_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449328512)))];
            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_to_fp16, x = var_523_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457717184)))];
            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_to_fp16, x = var_523_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(458765824)))];
            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_523_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_580_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_580_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_587_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_587_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_591_cast_fp16 = mul(x = x_1_cast_fp16, y = var_452_cast_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [2]> var_592_split_sizes_0 = const()[name = string("op_592_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_592_axis_0 = const()[name = string("op_592_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_592_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_592_cast_fp16_1 = split(axis = var_592_axis_0, split_sizes = var_592_split_sizes_0, x = x_1_cast_fp16)[name = string("op_592_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_594_cast_fp16 = mul(x = var_592_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_594_cast_fp16")];
            int32 var_596 = const()[name = string("op_596"), val = int32(-2)];
            bool var_597_interleave_0 = const()[name = string("op_597_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_597_cast_fp16 = concat(axis = var_596, interleave = var_597_interleave_0, values = (var_594_cast_fp16, var_592_cast_fp16_0))[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_598_cast_fp16 = mul(x = var_597_cast_fp16, y = var_459_cast_fp16)[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_3_cast_fp16 = add(x = var_591_cast_fp16, y = var_598_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_604_cast_fp16 = mul(x = var_580_cast_fp16, y = var_452_cast_fp16)[name = string("op_604_cast_fp16")];
            tensor<int32, [2]> var_605_split_sizes_0 = const()[name = string("op_605_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_605_axis_0 = const()[name = string("op_605_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_605_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_605_cast_fp16_1 = split(axis = var_605_axis_0, split_sizes = var_605_split_sizes_0, x = var_580_cast_fp16)[name = string("op_605_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_607_cast_fp16 = mul(x = var_605_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_607_cast_fp16")];
            int32 var_609 = const()[name = string("op_609"), val = int32(-2)];
            bool var_610_interleave_0 = const()[name = string("op_610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_610_cast_fp16 = concat(axis = var_609, interleave = var_610_interleave_0, values = (var_607_cast_fp16, var_605_cast_fp16_0))[name = string("op_610_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_611_cast_fp16 = mul(x = var_610_cast_fp16, y = var_459_cast_fp16)[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_5_cast_fp16 = add(x = var_604_cast_fp16, y = var_611_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 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_176")];
            tensor<fp16, [14, 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_84_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_84 = read_state(input = key_cache)[name = string("coreml_update_state_84")];
            tensor<fp16, [14, 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_587_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [14, 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_85_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_85 = read_state(input = value_cache)[name = string("coreml_update_state_85")];
            tensor<int32, [4]> var_681_begin_0 = const()[name = string("op_681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_681_end_0 = const()[name = string("op_681_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_681_end_mask_0 = const()[name = string("op_681_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_681_cast_fp16 = slice_by_index(begin = var_681_begin_0, end = var_681_end_0, end_mask = var_681_end_mask_0, x = coreml_update_state_84)[name = string("op_681_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_684_axis_0 = const()[name = string("op_684_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_1 = split(axis = var_684_axis_0, split_sizes = tile_0, x = var_681_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [4]> var_691_begin_0 = const()[name = string("op_691_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_691_end_0 = const()[name = string("op_691_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_691_end_mask_0 = const()[name = string("op_691_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = coreml_update_state_85)[name = string("op_691_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_694_axis_0 = const()[name = string("op_694_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_1 = split(axis = var_694_axis_0, split_sizes = tile_1, x = var_691_cast_fp16)[name = string("op_694_cast_fp16")];
            tensor<int32, [2]> var_697_split_sizes_0 = const()[name = string("op_697_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_697_axis_0 = const()[name = string("op_697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_697_0, tensor<fp16, [1, 8, 128, 16]> var_697_1 = split(axis = var_697_axis_0, split_sizes = var_697_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_697")];
            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_684_cast_fp16_0, y = var_697_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_700_to_fp16 = const()[name = string("op_700_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_700_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_704 = const()[name = string("op_704"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_7_cast_fp16 = softmax(axis = var_704, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_710_transpose_x_1 = const()[name = string("op_710_transpose_x_1"), val = bool(true)];
            bool var_710_transpose_y_1 = const()[name = string("op_710_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_710_cast_fp16 = matmul(transpose_x = var_710_transpose_x_1, transpose_y = var_710_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_694_cast_fp16_0)[name = string("op_710_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_684_cast_fp16_1, y = var_697_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_712_to_fp16 = const()[name = string("op_712_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_712_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_716 = const()[name = string("op_716"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_15_cast_fp16 = softmax(axis = var_716, 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_694_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_724 = const()[name = string("op_724"), 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_724, interleave = attn_output_3_interleave_0, values = (var_710_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_728_perm_0 = const()[name = string("op_728_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_728_cast_fp16 = transpose(perm = var_728_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_728_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(459814464)))];
            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_761_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_761_cast_fp16")];
            int32 var_759 = const()[name = string("op_759"), 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_759, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_761_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(468203136)))];
            fp16 var_771_to_fp16 = const()[name = string("op_771_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_771_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_782_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = out_3_cast_fp16)[name = string("op_782_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468211392)))];
            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_to_fp16, x = var_782_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_799_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_799_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493377280)))];
            tensor<int32, [2]> var_805_strides_0 = const()[name = string("op_805_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_805_pad_type_0 = const()[name = string("op_805_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_805_pad_0 = const()[name = string("op_805_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_805_dilations_0 = const()[name = string("op_805_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_805_groups_0 = const()[name = string("op_805_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_805_cast_fp16 = conv(dilations = var_805_dilations_0, groups = var_805_groups_0, pad = var_805_pad_0, pad_type = var_805_pad_type_0, strides = var_805_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_782_cast_fp16_0)[name = string("op_805_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_9_cast_fp16 = mul(x = var_799_cast_fp16, y = var_805_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518543168)))];
            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_to_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_823_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_823_cast_fp16")];
            int32 var_821 = const()[name = string("op_821"), 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_821, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_823_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(543709056)))];
            fp16 var_833_to_fp16 = const()[name = string("op_833_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_833_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_844_split_sizes_0 = const()[name = string("op_844_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_844_axis_0 = const()[name = string("op_844_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_844_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_844_cast_fp16_1 = split(axis = var_844_axis_0, split_sizes = var_844_split_sizes_0, x = out_5_cast_fp16)[name = string("op_844_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543717312)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552105984)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            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_cast_fp16, x = var_844_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_901_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_901_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_908_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_908_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_912_cast_fp16 = mul(x = x_11_cast_fp16, y = var_452_cast_fp16)[name = string("op_912_cast_fp16")];
            tensor<int32, [2]> var_913_split_sizes_0 = const()[name = string("op_913_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_913_axis_0 = const()[name = string("op_913_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_913_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_913_cast_fp16_1 = split(axis = var_913_axis_0, split_sizes = var_913_split_sizes_0, x = x_11_cast_fp16)[name = string("op_913_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_915_cast_fp16 = mul(x = var_913_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_915_cast_fp16")];
            int32 var_917 = const()[name = string("op_917"), val = int32(-2)];
            bool var_918_interleave_0 = const()[name = string("op_918_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_918_cast_fp16 = concat(axis = var_917, interleave = var_918_interleave_0, values = (var_915_cast_fp16, var_913_cast_fp16_0))[name = string("op_918_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_919_cast_fp16 = mul(x = var_918_cast_fp16, y = var_459_cast_fp16)[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_9_cast_fp16 = add(x = var_912_cast_fp16, y = var_919_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_925_cast_fp16 = mul(x = var_901_cast_fp16, y = var_452_cast_fp16)[name = string("op_925_cast_fp16")];
            tensor<int32, [2]> var_926_split_sizes_0 = const()[name = string("op_926_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_926_axis_0 = const()[name = string("op_926_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_926_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_926_cast_fp16_1 = split(axis = var_926_axis_0, split_sizes = var_926_split_sizes_0, x = var_901_cast_fp16)[name = string("op_926_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_928_cast_fp16 = mul(x = var_926_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_928_cast_fp16")];
            int32 var_930 = const()[name = string("op_930"), val = int32(-2)];
            bool var_931_interleave_0 = const()[name = string("op_931_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_931_cast_fp16 = concat(axis = var_930, interleave = var_931_interleave_0, values = (var_928_cast_fp16, var_926_cast_fp16_0))[name = string("op_931_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_932_cast_fp16 = mul(x = var_931_cast_fp16, y = var_459_cast_fp16)[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_15_cast_fp16 = add(x = var_925_cast_fp16, y = var_932_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_173")];
            tensor<fp16, [14, 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_84)[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_86_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_86 = read_state(input = key_cache)[name = string("coreml_update_state_86")];
            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_908_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [14, 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_85)[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_87_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_87 = read_state(input = value_cache)[name = string("coreml_update_state_87")];
            tensor<int32, [4]> var_1002_begin_0 = const()[name = string("op_1002_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1002_end_0 = const()[name = string("op_1002_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1002_end_mask_0 = const()[name = string("op_1002_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1002_cast_fp16 = slice_by_index(begin = var_1002_begin_0, end = var_1002_end_0, end_mask = var_1002_end_mask_0, x = coreml_update_state_86)[name = string("op_1002_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1005_axis_0 = const()[name = string("op_1005_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_1 = split(axis = var_1005_axis_0, split_sizes = tile_2, x = var_1002_cast_fp16)[name = string("op_1005_cast_fp16")];
            tensor<int32, [4]> var_1012_begin_0 = const()[name = string("op_1012_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1012_end_0 = const()[name = string("op_1012_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1012_end_mask_0 = const()[name = string("op_1012_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1012_cast_fp16 = slice_by_index(begin = var_1012_begin_0, end = var_1012_end_0, end_mask = var_1012_end_mask_0, x = coreml_update_state_87)[name = string("op_1012_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1015_axis_0 = const()[name = string("op_1015_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_1 = split(axis = var_1015_axis_0, split_sizes = tile_3, x = var_1012_cast_fp16)[name = string("op_1015_cast_fp16")];
            tensor<int32, [2]> var_1018_split_sizes_0 = const()[name = string("op_1018_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1018_axis_0 = const()[name = string("op_1018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1018_0, tensor<fp16, [1, 8, 128, 16]> var_1018_1 = split(axis = var_1018_axis_0, split_sizes = var_1018_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1018")];
            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_1005_cast_fp16_0, y = var_1018_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1021_to_fp16 = const()[name = string("op_1021_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_1021_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_1025 = const()[name = string("op_1025"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_23_cast_fp16 = softmax(axis = var_1025, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1031_transpose_x_1 = const()[name = string("op_1031_transpose_x_1"), val = bool(true)];
            bool var_1031_transpose_y_1 = const()[name = string("op_1031_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1031_cast_fp16 = matmul(transpose_x = var_1031_transpose_x_1, transpose_y = var_1031_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1015_cast_fp16_0)[name = string("op_1031_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_1005_cast_fp16_1, y = var_1018_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1033_to_fp16 = const()[name = string("op_1033_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_1033_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_1037 = const()[name = string("op_1037"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_31_cast_fp16 = softmax(axis = var_1037, 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_1015_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1045 = const()[name = string("op_1045"), 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_1045, interleave = attn_output_11_interleave_0, values = (var_1031_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1049_perm_0 = const()[name = string("op_1049_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_1049_cast_fp16 = transpose(perm = var_1049_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1049_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553154624)))];
            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_to_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_1082_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1082_cast_fp16")];
            int32 var_1080 = const()[name = string("op_1080"), 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_1080, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1082_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(561543296)))];
            fp16 var_1092_to_fp16 = const()[name = string("op_1092_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_1092_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1103_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1103_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561551552)))];
            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_to_fp16, x = var_1103_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_1120_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [2]> var_1126_strides_0 = const()[name = string("op_1126_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1126_pad_type_0 = const()[name = string("op_1126_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1126_pad_0 = const()[name = string("op_1126_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1126_dilations_0 = const()[name = string("op_1126_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1126_groups_0 = const()[name = string("op_1126_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_1126_cast_fp16 = conv(dilations = var_1126_dilations_0, groups = var_1126_groups_0, pad = var_1126_pad_0, pad_type = var_1126_pad_type_0, strides = var_1126_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1103_cast_fp16_0)[name = string("op_1126_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_19_cast_fp16 = mul(x = var_1120_cast_fp16, y = var_1126_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586717440)))];
            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_to_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_1144_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1144_cast_fp16")];
            int32 var_1142 = const()[name = string("op_1142"), 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_1142, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1144_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(611883328)))];
            fp16 var_1154_to_fp16 = const()[name = string("op_1154_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_1154_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1165_split_sizes_0 = const()[name = string("op_1165_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1165_axis_0 = const()[name = string("op_1165_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1165_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1165_cast_fp16_1 = split(axis = var_1165_axis_0, split_sizes = var_1165_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1165_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611891584)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620280256)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            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_cast_fp16, x = var_1165_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_1222_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1222_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_1229_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1229_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1233_cast_fp16 = mul(x = x_21_cast_fp16, y = var_452_cast_fp16)[name = string("op_1233_cast_fp16")];
            tensor<int32, [2]> var_1234_split_sizes_0 = const()[name = string("op_1234_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1234_axis_0 = const()[name = string("op_1234_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_1234_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_1234_cast_fp16_1 = split(axis = var_1234_axis_0, split_sizes = var_1234_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1234_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_1236_cast_fp16 = mul(x = var_1234_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1236_cast_fp16")];
            int32 var_1238 = const()[name = string("op_1238"), val = int32(-2)];
            bool var_1239_interleave_0 = const()[name = string("op_1239_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_1239_cast_fp16 = concat(axis = var_1238, interleave = var_1239_interleave_0, values = (var_1236_cast_fp16, var_1234_cast_fp16_0))[name = string("op_1239_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1240_cast_fp16 = mul(x = var_1239_cast_fp16, y = var_459_cast_fp16)[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_15_cast_fp16 = add(x = var_1233_cast_fp16, y = var_1240_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1246_cast_fp16 = mul(x = var_1222_cast_fp16, y = var_452_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<int32, [2]> var_1247_split_sizes_0 = const()[name = string("op_1247_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1247_axis_0 = const()[name = string("op_1247_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_1247_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_1247_cast_fp16_1 = split(axis = var_1247_axis_0, split_sizes = var_1247_split_sizes_0, x = var_1222_cast_fp16)[name = string("op_1247_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_1249_cast_fp16 = mul(x = var_1247_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1249_cast_fp16")];
            int32 var_1251 = const()[name = string("op_1251"), val = int32(-2)];
            bool var_1252_interleave_0 = const()[name = string("op_1252_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_1252_cast_fp16 = concat(axis = var_1251, interleave = var_1252_interleave_0, values = (var_1249_cast_fp16, var_1247_cast_fp16_0))[name = string("op_1252_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1253_cast_fp16 = mul(x = var_1252_cast_fp16, y = var_459_cast_fp16)[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_25_cast_fp16 = add(x = var_1246_cast_fp16, y = var_1253_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_170")];
            tensor<fp16, [14, 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_86)[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_88_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_88 = read_state(input = key_cache)[name = string("coreml_update_state_88")];
            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_1229_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [14, 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_87)[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_89_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_89 = read_state(input = value_cache)[name = string("coreml_update_state_89")];
            tensor<int32, [4]> var_1323_begin_0 = const()[name = string("op_1323_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1323_end_0 = const()[name = string("op_1323_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1323_end_mask_0 = const()[name = string("op_1323_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1323_cast_fp16 = slice_by_index(begin = var_1323_begin_0, end = var_1323_end_0, end_mask = var_1323_end_mask_0, x = coreml_update_state_88)[name = string("op_1323_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1326_axis_0 = const()[name = string("op_1326_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_1 = split(axis = var_1326_axis_0, split_sizes = tile_4, x = var_1323_cast_fp16)[name = string("op_1326_cast_fp16")];
            tensor<int32, [4]> var_1333_begin_0 = const()[name = string("op_1333_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1333_end_0 = const()[name = string("op_1333_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1333_end_mask_0 = const()[name = string("op_1333_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1333_cast_fp16 = slice_by_index(begin = var_1333_begin_0, end = var_1333_end_0, end_mask = var_1333_end_mask_0, x = coreml_update_state_89)[name = string("op_1333_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1336_axis_0 = const()[name = string("op_1336_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_1 = split(axis = var_1336_axis_0, split_sizes = tile_5, x = var_1333_cast_fp16)[name = string("op_1336_cast_fp16")];
            tensor<int32, [2]> var_1339_split_sizes_0 = const()[name = string("op_1339_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1339_axis_0 = const()[name = string("op_1339_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1339_0, tensor<fp16, [1, 8, 128, 16]> var_1339_1 = split(axis = var_1339_axis_0, split_sizes = var_1339_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1339")];
            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_1326_cast_fp16_0, y = var_1339_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1342_to_fp16 = const()[name = string("op_1342_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_1342_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_1346 = const()[name = string("op_1346"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_39_cast_fp16 = softmax(axis = var_1346, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1352_transpose_x_1 = const()[name = string("op_1352_transpose_x_1"), val = bool(true)];
            bool var_1352_transpose_y_1 = const()[name = string("op_1352_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1352_cast_fp16 = matmul(transpose_x = var_1352_transpose_x_1, transpose_y = var_1352_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1336_cast_fp16_0)[name = string("op_1352_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_1326_cast_fp16_1, y = var_1339_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1354_to_fp16 = const()[name = string("op_1354_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_1354_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_1358 = const()[name = string("op_1358"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_47_cast_fp16 = softmax(axis = var_1358, 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_1336_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1366 = const()[name = string("op_1366"), 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_1366, interleave = attn_output_19_interleave_0, values = (var_1352_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1370_perm_0 = const()[name = string("op_1370_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_1370_cast_fp16 = transpose(perm = var_1370_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1370_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621328896)))];
            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_to_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_1403_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1403_cast_fp16")];
            int32 var_1401 = const()[name = string("op_1401"), 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_1401, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1403_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(629717568)))];
            fp16 var_1413_to_fp16 = const()[name = string("op_1413_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_1413_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1424_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1424_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629725824)))];
            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_to_fp16, x = var_1424_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_1441_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1441_cast_fp16")];
            tensor<int32, [2]> var_1447_strides_0 = const()[name = string("op_1447_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1447_pad_type_0 = const()[name = string("op_1447_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1447_pad_0 = const()[name = string("op_1447_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1447_dilations_0 = const()[name = string("op_1447_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1447_groups_0 = const()[name = string("op_1447_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_1447_cast_fp16 = conv(dilations = var_1447_dilations_0, groups = var_1447_groups_0, pad = var_1447_pad_0, pad_type = var_1447_pad_type_0, strides = var_1447_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1424_cast_fp16_0)[name = string("op_1447_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_29_cast_fp16 = mul(x = var_1441_cast_fp16, y = var_1447_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654891712)))];
            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_to_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_1465_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1465_cast_fp16")];
            int32 var_1463 = const()[name = string("op_1463"), 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_1463, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1465_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(680057600)))];
            fp16 var_1475_to_fp16 = const()[name = string("op_1475_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_1475_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1486_split_sizes_0 = const()[name = string("op_1486_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1486_axis_0 = const()[name = string("op_1486_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1486_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1486_cast_fp16_1 = split(axis = var_1486_axis_0, split_sizes = var_1486_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1486_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680065856)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688454528)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            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_cast_fp16, x = var_1486_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_1543_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1543_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_1550_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1554_cast_fp16 = mul(x = x_31_cast_fp16, y = var_452_cast_fp16)[name = string("op_1554_cast_fp16")];
            tensor<int32, [2]> var_1555_split_sizes_0 = const()[name = string("op_1555_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1555_axis_0 = const()[name = string("op_1555_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_1555_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_1555_cast_fp16_1 = split(axis = var_1555_axis_0, split_sizes = var_1555_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1555_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_1557_cast_fp16 = mul(x = var_1555_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1557_cast_fp16")];
            int32 var_1559 = const()[name = string("op_1559"), val = int32(-2)];
            bool var_1560_interleave_0 = const()[name = string("op_1560_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_1560_cast_fp16 = concat(axis = var_1559, interleave = var_1560_interleave_0, values = (var_1557_cast_fp16, var_1555_cast_fp16_0))[name = string("op_1560_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1561_cast_fp16 = mul(x = var_1560_cast_fp16, y = var_459_cast_fp16)[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_21_cast_fp16 = add(x = var_1554_cast_fp16, y = var_1561_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1567_cast_fp16 = mul(x = var_1543_cast_fp16, y = var_452_cast_fp16)[name = string("op_1567_cast_fp16")];
            tensor<int32, [2]> var_1568_split_sizes_0 = const()[name = string("op_1568_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1568_axis_0 = const()[name = string("op_1568_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_1568_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_1568_cast_fp16_1 = split(axis = var_1568_axis_0, split_sizes = var_1568_split_sizes_0, x = var_1543_cast_fp16)[name = string("op_1568_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_1570_cast_fp16 = mul(x = var_1568_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1570_cast_fp16")];
            int32 var_1572 = const()[name = string("op_1572"), val = int32(-2)];
            bool var_1573_interleave_0 = const()[name = string("op_1573_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_1573_cast_fp16 = concat(axis = var_1572, interleave = var_1573_interleave_0, values = (var_1570_cast_fp16, var_1568_cast_fp16_0))[name = string("op_1573_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1574_cast_fp16 = mul(x = var_1573_cast_fp16, y = var_459_cast_fp16)[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_35_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1574_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_167")];
            tensor<fp16, [14, 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_88)[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_90_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_90 = read_state(input = key_cache)[name = string("coreml_update_state_90")];
            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_1550_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [14, 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_89)[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_91_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_91 = read_state(input = value_cache)[name = string("coreml_update_state_91")];
            tensor<int32, [4]> var_1644_begin_0 = const()[name = string("op_1644_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1644_end_0 = const()[name = string("op_1644_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1644_end_mask_0 = const()[name = string("op_1644_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1644_cast_fp16 = slice_by_index(begin = var_1644_begin_0, end = var_1644_end_0, end_mask = var_1644_end_mask_0, x = coreml_update_state_90)[name = string("op_1644_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1647_axis_0 = const()[name = string("op_1647_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_1 = split(axis = var_1647_axis_0, split_sizes = tile_6, x = var_1644_cast_fp16)[name = string("op_1647_cast_fp16")];
            tensor<int32, [4]> var_1654_begin_0 = const()[name = string("op_1654_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1654_end_0 = const()[name = string("op_1654_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1654_end_mask_0 = const()[name = string("op_1654_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1654_cast_fp16 = slice_by_index(begin = var_1654_begin_0, end = var_1654_end_0, end_mask = var_1654_end_mask_0, x = coreml_update_state_91)[name = string("op_1654_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1657_axis_0 = const()[name = string("op_1657_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_1 = split(axis = var_1657_axis_0, split_sizes = tile_7, x = var_1654_cast_fp16)[name = string("op_1657_cast_fp16")];
            tensor<int32, [2]> var_1660_split_sizes_0 = const()[name = string("op_1660_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1660_axis_0 = const()[name = string("op_1660_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1660_0, tensor<fp16, [1, 8, 128, 16]> var_1660_1 = split(axis = var_1660_axis_0, split_sizes = var_1660_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1660")];
            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_1647_cast_fp16_0, y = var_1660_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1663_to_fp16 = const()[name = string("op_1663_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_1663_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_1667 = const()[name = string("op_1667"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_55_cast_fp16 = softmax(axis = var_1667, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1673_transpose_x_1 = const()[name = string("op_1673_transpose_x_1"), val = bool(true)];
            bool var_1673_transpose_y_1 = const()[name = string("op_1673_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1673_cast_fp16 = matmul(transpose_x = var_1673_transpose_x_1, transpose_y = var_1673_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1657_cast_fp16_0)[name = string("op_1673_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_1647_cast_fp16_1, y = var_1660_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1675_to_fp16 = const()[name = string("op_1675_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_1675_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_1679 = const()[name = string("op_1679"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_63_cast_fp16 = softmax(axis = var_1679, 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_1657_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1687 = const()[name = string("op_1687"), 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_1687, interleave = attn_output_27_interleave_0, values = (var_1673_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1691_perm_0 = const()[name = string("op_1691_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_1691_cast_fp16 = transpose(perm = var_1691_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1691_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            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_cast_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_1724_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1724_cast_fp16")];
            int32 var_1722 = const()[name = string("op_1722"), 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_1722, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1724_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(689503168)))];
            fp16 var_1734_to_fp16 = const()[name = string("op_1734_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_1734_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1745_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1745_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689511424)))];
            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_to_fp16, x = var_1745_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_1762_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1762_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(714677312)))];
            tensor<int32, [2]> var_1768_strides_0 = const()[name = string("op_1768_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1768_pad_type_0 = const()[name = string("op_1768_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1768_pad_0 = const()[name = string("op_1768_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1768_dilations_0 = const()[name = string("op_1768_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1768_groups_0 = const()[name = string("op_1768_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_1768_cast_fp16 = conv(dilations = var_1768_dilations_0, groups = var_1768_groups_0, pad = var_1768_pad_0, pad_type = var_1768_pad_type_0, strides = var_1768_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1745_cast_fp16_0)[name = string("op_1768_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_39_cast_fp16 = mul(x = var_1762_cast_fp16, y = var_1768_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_1786_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1786_cast_fp16")];
            int32 var_1784 = const()[name = string("op_1784"), 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_1784, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1786_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(739843200)))];
            fp16 var_1796_to_fp16 = const()[name = string("op_1796_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_1796_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1807_split_sizes_0 = const()[name = string("op_1807_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1807_axis_0 = const()[name = string("op_1807_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_1807_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_1807_cast_fp16_1 = split(axis = var_1807_axis_0, split_sizes = var_1807_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739851456)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(748240128)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            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_cast_fp16, x = var_1807_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_1864_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1864_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_1871_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1871_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1875_cast_fp16 = mul(x = x_41_cast_fp16, y = var_452_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [2]> var_1876_split_sizes_0 = const()[name = string("op_1876_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1876_axis_0 = const()[name = string("op_1876_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_1876_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_1876_cast_fp16_1 = split(axis = var_1876_axis_0, split_sizes = var_1876_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1876_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_1878_cast_fp16 = mul(x = var_1876_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1878_cast_fp16")];
            int32 var_1880 = const()[name = string("op_1880"), val = int32(-2)];
            bool var_1881_interleave_0 = const()[name = string("op_1881_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_1881_cast_fp16 = concat(axis = var_1880, interleave = var_1881_interleave_0, values = (var_1878_cast_fp16, var_1876_cast_fp16_0))[name = string("op_1881_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_1882_cast_fp16 = mul(x = var_1881_cast_fp16, y = var_459_cast_fp16)[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_27_cast_fp16 = add(x = var_1875_cast_fp16, y = var_1882_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1888_cast_fp16 = mul(x = var_1864_cast_fp16, y = var_452_cast_fp16)[name = string("op_1888_cast_fp16")];
            tensor<int32, [2]> var_1889_split_sizes_0 = const()[name = string("op_1889_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1889_axis_0 = const()[name = string("op_1889_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_1889_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_1889_cast_fp16_1 = split(axis = var_1889_axis_0, split_sizes = var_1889_split_sizes_0, x = var_1864_cast_fp16)[name = string("op_1889_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_1891_cast_fp16 = mul(x = var_1889_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1891_cast_fp16")];
            int32 var_1893 = const()[name = string("op_1893"), val = int32(-2)];
            bool var_1894_interleave_0 = const()[name = string("op_1894_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_1894_cast_fp16 = concat(axis = var_1893, interleave = var_1894_interleave_0, values = (var_1891_cast_fp16, var_1889_cast_fp16_0))[name = string("op_1894_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_1895_cast_fp16 = mul(x = var_1894_cast_fp16, y = var_459_cast_fp16)[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_45_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1895_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_164")];
            tensor<fp16, [14, 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_90)[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_92_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_92 = read_state(input = key_cache)[name = string("coreml_update_state_92")];
            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_1871_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [14, 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_91)[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_93_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_93 = read_state(input = value_cache)[name = string("coreml_update_state_93")];
            tensor<int32, [4]> var_1965_begin_0 = const()[name = string("op_1965_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1965_end_0 = const()[name = string("op_1965_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1965_end_mask_0 = const()[name = string("op_1965_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1965_cast_fp16 = slice_by_index(begin = var_1965_begin_0, end = var_1965_end_0, end_mask = var_1965_end_mask_0, x = coreml_update_state_92)[name = string("op_1965_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1968_axis_0 = const()[name = string("op_1968_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_1 = split(axis = var_1968_axis_0, split_sizes = tile_8, x = var_1965_cast_fp16)[name = string("op_1968_cast_fp16")];
            tensor<int32, [4]> var_1975_begin_0 = const()[name = string("op_1975_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1975_end_0 = const()[name = string("op_1975_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1975_end_mask_0 = const()[name = string("op_1975_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1975_cast_fp16 = slice_by_index(begin = var_1975_begin_0, end = var_1975_end_0, end_mask = var_1975_end_mask_0, x = coreml_update_state_93)[name = string("op_1975_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1978_axis_0 = const()[name = string("op_1978_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_1 = split(axis = var_1978_axis_0, split_sizes = tile_9, x = var_1975_cast_fp16)[name = string("op_1978_cast_fp16")];
            tensor<int32, [2]> var_1981_split_sizes_0 = const()[name = string("op_1981_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1981_axis_0 = const()[name = string("op_1981_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_1981_0, tensor<fp16, [1, 8, 128, 16]> var_1981_1 = split(axis = var_1981_axis_0, split_sizes = var_1981_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1981")];
            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_1968_cast_fp16_0, y = var_1981_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1984_to_fp16 = const()[name = string("op_1984_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_1984_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_1988 = const()[name = string("op_1988"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_71_cast_fp16 = softmax(axis = var_1988, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1994_transpose_x_1 = const()[name = string("op_1994_transpose_x_1"), val = bool(true)];
            bool var_1994_transpose_y_1 = const()[name = string("op_1994_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_1994_cast_fp16 = matmul(transpose_x = var_1994_transpose_x_1, transpose_y = var_1994_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1978_cast_fp16_0)[name = string("op_1994_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_1968_cast_fp16_1, y = var_1981_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1996_to_fp16 = const()[name = string("op_1996_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_1996_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_2000 = const()[name = string("op_2000"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_79_cast_fp16 = softmax(axis = var_2000, 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_1978_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2008 = const()[name = string("op_2008"), 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_2008, interleave = attn_output_35_interleave_0, values = (var_1994_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2012_perm_0 = const()[name = string("op_2012_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_2012_cast_fp16 = transpose(perm = var_2012_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2012_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_2045_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2045_cast_fp16")];
            int32 var_2043 = const()[name = string("op_2043"), 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_2043, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2045_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(749288768)))];
            fp16 var_2055_to_fp16 = const()[name = string("op_2055_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_2055_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2066_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2066_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_2066_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_2083_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2083_cast_fp16")];
            tensor<int32, [2]> var_2089_strides_0 = const()[name = string("op_2089_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2089_pad_type_0 = const()[name = string("op_2089_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2089_pad_0 = const()[name = string("op_2089_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2089_dilations_0 = const()[name = string("op_2089_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2089_groups_0 = const()[name = string("op_2089_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_2089_cast_fp16 = conv(dilations = var_2089_dilations_0, groups = var_2089_groups_0, pad = var_2089_pad_0, pad_type = var_2089_pad_type_0, strides = var_2089_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2066_cast_fp16_0)[name = string("op_2089_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_49_cast_fp16 = mul(x = var_2083_cast_fp16, y = var_2089_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_2107_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2107_cast_fp16")];
            int32 var_2105 = const()[name = string("op_2105"), 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_2105, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2107_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(749297024)))];
            fp16 var_2117_to_fp16 = const()[name = string("op_2117_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_2117_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2128_split_sizes_0 = const()[name = string("op_2128_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2128_axis_0 = const()[name = string("op_2128_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2128_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2128_cast_fp16_1 = split(axis = var_2128_axis_0, split_sizes = var_2128_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749305280)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(757693952)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            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_cast_fp16, x = var_2128_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_2185_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2185_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_2192_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2196_cast_fp16 = mul(x = x_51_cast_fp16, y = var_452_cast_fp16)[name = string("op_2196_cast_fp16")];
            tensor<int32, [2]> var_2197_split_sizes_0 = const()[name = string("op_2197_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2197_axis_0 = const()[name = string("op_2197_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_2197_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_2197_cast_fp16_1 = split(axis = var_2197_axis_0, split_sizes = var_2197_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2197_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_2199_cast_fp16 = mul(x = var_2197_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2199_cast_fp16")];
            int32 var_2201 = const()[name = string("op_2201"), val = int32(-2)];
            bool var_2202_interleave_0 = const()[name = string("op_2202_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_2202_cast_fp16 = concat(axis = var_2201, interleave = var_2202_interleave_0, values = (var_2199_cast_fp16, var_2197_cast_fp16_0))[name = string("op_2202_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2203_cast_fp16 = mul(x = var_2202_cast_fp16, y = var_459_cast_fp16)[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_33_cast_fp16 = add(x = var_2196_cast_fp16, y = var_2203_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2209_cast_fp16 = mul(x = var_2185_cast_fp16, y = var_452_cast_fp16)[name = string("op_2209_cast_fp16")];
            tensor<int32, [2]> var_2210_split_sizes_0 = const()[name = string("op_2210_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2210_axis_0 = const()[name = string("op_2210_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_2210_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_2210_cast_fp16_1 = split(axis = var_2210_axis_0, split_sizes = var_2210_split_sizes_0, x = var_2185_cast_fp16)[name = string("op_2210_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_2212_cast_fp16 = mul(x = var_2210_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2212_cast_fp16")];
            int32 var_2214 = const()[name = string("op_2214"), val = int32(-2)];
            bool var_2215_interleave_0 = const()[name = string("op_2215_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_2215_cast_fp16 = concat(axis = var_2214, interleave = var_2215_interleave_0, values = (var_2212_cast_fp16, var_2210_cast_fp16_0))[name = string("op_2215_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2216_cast_fp16 = mul(x = var_2215_cast_fp16, y = var_459_cast_fp16)[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_55_cast_fp16 = add(x = var_2209_cast_fp16, y = var_2216_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_161")];
            tensor<fp16, [14, 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_92)[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_94_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_94 = read_state(input = key_cache)[name = string("coreml_update_state_94")];
            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_2192_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [14, 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_93)[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_95_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_95 = read_state(input = value_cache)[name = string("coreml_update_state_95")];
            tensor<int32, [4]> var_2286_begin_0 = const()[name = string("op_2286_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2286_end_0 = const()[name = string("op_2286_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2286_end_mask_0 = const()[name = string("op_2286_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2286_cast_fp16 = slice_by_index(begin = var_2286_begin_0, end = var_2286_end_0, end_mask = var_2286_end_mask_0, x = coreml_update_state_94)[name = string("op_2286_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2289_axis_0 = const()[name = string("op_2289_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_1 = split(axis = var_2289_axis_0, split_sizes = tile_10, x = var_2286_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> var_2296_begin_0 = const()[name = string("op_2296_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2296_end_0 = const()[name = string("op_2296_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2296_end_mask_0 = const()[name = string("op_2296_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2296_cast_fp16 = slice_by_index(begin = var_2296_begin_0, end = var_2296_end_0, end_mask = var_2296_end_mask_0, x = coreml_update_state_95)[name = string("op_2296_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2299_axis_0 = const()[name = string("op_2299_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_1 = split(axis = var_2299_axis_0, split_sizes = tile_11, x = var_2296_cast_fp16)[name = string("op_2299_cast_fp16")];
            tensor<int32, [2]> var_2302_split_sizes_0 = const()[name = string("op_2302_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2302_axis_0 = const()[name = string("op_2302_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_2302_0, tensor<fp16, [1, 8, 128, 16]> var_2302_1 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2302")];
            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_2289_cast_fp16_0, y = var_2302_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2305_to_fp16 = const()[name = string("op_2305_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_2305_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_2309 = const()[name = string("op_2309"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_87_cast_fp16 = softmax(axis = var_2309, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2315_transpose_x_1 = const()[name = string("op_2315_transpose_x_1"), val = bool(true)];
            bool var_2315_transpose_y_1 = const()[name = string("op_2315_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_2315_cast_fp16 = matmul(transpose_x = var_2315_transpose_x_1, transpose_y = var_2315_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2299_cast_fp16_0)[name = string("op_2315_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_2289_cast_fp16_1, y = var_2302_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2317_to_fp16 = const()[name = string("op_2317_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_2317_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_2321 = const()[name = string("op_2321"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_95_cast_fp16 = softmax(axis = var_2321, 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_2299_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2329 = const()[name = string("op_2329"), 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_2329, interleave = attn_output_43_interleave_0, values = (var_2315_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2333_perm_0 = const()[name = string("op_2333_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_2333_cast_fp16 = transpose(perm = var_2333_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2333_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_2366_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2366_cast_fp16")];
            int32 var_2364 = const()[name = string("op_2364"), 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_2364, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2366_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(758742592)))];
            fp16 var_2376_to_fp16 = const()[name = string("op_2376_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_2376_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2387_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2387_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(758750848)))];
            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_to_fp16, x = var_2387_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_2404_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [2]> var_2410_strides_0 = const()[name = string("op_2410_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2410_pad_type_0 = const()[name = string("op_2410_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2410_pad_0 = const()[name = string("op_2410_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2410_dilations_0 = const()[name = string("op_2410_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2410_groups_0 = const()[name = string("op_2410_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_2410_cast_fp16 = conv(dilations = var_2410_dilations_0, groups = var_2410_groups_0, pad = var_2410_pad_0, pad_type = var_2410_pad_type_0, strides = var_2410_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2387_cast_fp16_0)[name = string("op_2410_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_59_cast_fp16 = mul(x = var_2404_cast_fp16, y = var_2410_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_2428_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2428_cast_fp16")];
            int32 var_2426 = const()[name = string("op_2426"), 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_2426, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2428_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(783916736)))];
            fp16 var_2438_to_fp16 = const()[name = string("op_2438_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_2438_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2449_split_sizes_0 = const()[name = string("op_2449_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2449_axis_0 = const()[name = string("op_2449_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2449_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2449_cast_fp16_1 = split(axis = var_2449_axis_0, split_sizes = var_2449_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2449_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(783924992)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792313664)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            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_cast_fp16, x = var_2449_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_2506_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2506_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_2513_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2517_cast_fp16 = mul(x = x_61_cast_fp16, y = var_452_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<int32, [2]> var_2518_split_sizes_0 = const()[name = string("op_2518_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2518_axis_0 = const()[name = string("op_2518_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_2518_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_2518_cast_fp16_1 = split(axis = var_2518_axis_0, split_sizes = var_2518_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2518_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_2520_cast_fp16 = mul(x = var_2518_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2520_cast_fp16")];
            int32 var_2522 = const()[name = string("op_2522"), val = int32(-2)];
            bool var_2523_interleave_0 = const()[name = string("op_2523_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_2523_cast_fp16 = concat(axis = var_2522, interleave = var_2523_interleave_0, values = (var_2520_cast_fp16, var_2518_cast_fp16_0))[name = string("op_2523_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2524_cast_fp16 = mul(x = var_2523_cast_fp16, y = var_459_cast_fp16)[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_39_cast_fp16 = add(x = var_2517_cast_fp16, y = var_2524_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2530_cast_fp16 = mul(x = var_2506_cast_fp16, y = var_452_cast_fp16)[name = string("op_2530_cast_fp16")];
            tensor<int32, [2]> var_2531_split_sizes_0 = const()[name = string("op_2531_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2531_axis_0 = const()[name = string("op_2531_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_2531_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_2531_cast_fp16_1 = split(axis = var_2531_axis_0, split_sizes = var_2531_split_sizes_0, x = var_2506_cast_fp16)[name = string("op_2531_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_2533_cast_fp16 = mul(x = var_2531_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2533_cast_fp16")];
            int32 var_2535 = const()[name = string("op_2535"), val = int32(-2)];
            bool var_2536_interleave_0 = const()[name = string("op_2536_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_2536_cast_fp16 = concat(axis = var_2535, interleave = var_2536_interleave_0, values = (var_2533_cast_fp16, var_2531_cast_fp16_0))[name = string("op_2536_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2537_cast_fp16 = mul(x = var_2536_cast_fp16, y = var_459_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_65_cast_fp16 = add(x = var_2530_cast_fp16, y = var_2537_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_158")];
            tensor<fp16, [14, 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_94)[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_96_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_96 = read_state(input = key_cache)[name = string("coreml_update_state_96")];
            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_2513_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [14, 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_95)[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_97_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_97 = read_state(input = value_cache)[name = string("coreml_update_state_97")];
            tensor<int32, [4]> var_2607_begin_0 = const()[name = string("op_2607_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2607_end_0 = const()[name = string("op_2607_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2607_end_mask_0 = const()[name = string("op_2607_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2607_cast_fp16 = slice_by_index(begin = var_2607_begin_0, end = var_2607_end_0, end_mask = var_2607_end_mask_0, x = coreml_update_state_96)[name = string("op_2607_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2610_axis_0 = const()[name = string("op_2610_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_1 = split(axis = var_2610_axis_0, split_sizes = tile_12, x = var_2607_cast_fp16)[name = string("op_2610_cast_fp16")];
            tensor<int32, [4]> var_2617_begin_0 = const()[name = string("op_2617_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2617_end_0 = const()[name = string("op_2617_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2617_end_mask_0 = const()[name = string("op_2617_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2617_cast_fp16 = slice_by_index(begin = var_2617_begin_0, end = var_2617_end_0, end_mask = var_2617_end_mask_0, x = coreml_update_state_97)[name = string("op_2617_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2620_axis_0 = const()[name = string("op_2620_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_1 = split(axis = var_2620_axis_0, split_sizes = tile_13, x = var_2617_cast_fp16)[name = string("op_2620_cast_fp16")];
            tensor<int32, [2]> var_2623_split_sizes_0 = const()[name = string("op_2623_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2623_axis_0 = const()[name = string("op_2623_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_2623_0, tensor<fp16, [1, 8, 128, 16]> var_2623_1 = split(axis = var_2623_axis_0, split_sizes = var_2623_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2623")];
            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_2610_cast_fp16_0, y = var_2623_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2626_to_fp16 = const()[name = string("op_2626_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_2626_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_2630 = const()[name = string("op_2630"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_103_cast_fp16 = softmax(axis = var_2630, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2636_transpose_x_1 = const()[name = string("op_2636_transpose_x_1"), val = bool(true)];
            bool var_2636_transpose_y_1 = const()[name = string("op_2636_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_2636_cast_fp16 = matmul(transpose_x = var_2636_transpose_x_1, transpose_y = var_2636_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2620_cast_fp16_0)[name = string("op_2636_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_2610_cast_fp16_1, y = var_2623_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2638_to_fp16 = const()[name = string("op_2638_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_2638_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_2642 = const()[name = string("op_2642"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_111_cast_fp16 = softmax(axis = var_2642, 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_2620_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2650 = const()[name = string("op_2650"), 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_2650, interleave = attn_output_51_interleave_0, values = (var_2636_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2654_perm_0 = const()[name = string("op_2654_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_2654_cast_fp16 = transpose(perm = var_2654_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2654_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_2687_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2687_cast_fp16")];
            int32 var_2685 = const()[name = string("op_2685"), 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_2685, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2687_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(793362304)))];
            fp16 var_2697_to_fp16 = const()[name = string("op_2697_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_2697_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2708_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2708_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_2708_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_2725_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2725_cast_fp16")];
            tensor<int32, [2]> var_2731_strides_0 = const()[name = string("op_2731_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2731_pad_type_0 = const()[name = string("op_2731_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2731_pad_0 = const()[name = string("op_2731_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2731_dilations_0 = const()[name = string("op_2731_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2731_groups_0 = const()[name = string("op_2731_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_2731_cast_fp16 = conv(dilations = var_2731_dilations_0, groups = var_2731_groups_0, pad = var_2731_pad_0, pad_type = var_2731_pad_type_0, strides = var_2731_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2708_cast_fp16_0)[name = string("op_2731_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_69_cast_fp16 = mul(x = var_2725_cast_fp16, y = var_2731_cast_fp16)[name = string("x_69_cast_fp16")];
            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_cast_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_2749_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2749_cast_fp16")];
            int32 var_2747 = const()[name = string("op_2747"), 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_2747, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2749_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(793370560)))];
            fp16 var_2759_to_fp16 = const()[name = string("op_2759_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_2759_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2770_split_sizes_0 = const()[name = string("op_2770_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2770_axis_0 = const()[name = string("op_2770_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_2770_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_2770_cast_fp16_1 = split(axis = var_2770_axis_0, split_sizes = var_2770_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2770_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(793378816)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(801767488)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            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_cast_fp16, x = var_2770_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_2827_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2827_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_2834_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2834_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2838_cast_fp16 = mul(x = x_71_cast_fp16, y = var_452_cast_fp16)[name = string("op_2838_cast_fp16")];
            tensor<int32, [2]> var_2839_split_sizes_0 = const()[name = string("op_2839_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2839_axis_0 = const()[name = string("op_2839_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_2839_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_2839_cast_fp16_1 = split(axis = var_2839_axis_0, split_sizes = var_2839_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2839_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_2841_cast_fp16 = mul(x = var_2839_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2841_cast_fp16")];
            int32 var_2843 = const()[name = string("op_2843"), val = int32(-2)];
            bool var_2844_interleave_0 = const()[name = string("op_2844_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_2844_cast_fp16 = concat(axis = var_2843, interleave = var_2844_interleave_0, values = (var_2841_cast_fp16, var_2839_cast_fp16_0))[name = string("op_2844_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_2845_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_459_cast_fp16)[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_45_cast_fp16 = add(x = var_2838_cast_fp16, y = var_2845_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2851_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_452_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [2]> var_2852_split_sizes_0 = const()[name = string("op_2852_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2852_axis_0 = const()[name = string("op_2852_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_2852_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_2852_cast_fp16_1 = split(axis = var_2852_axis_0, split_sizes = var_2852_split_sizes_0, x = var_2827_cast_fp16)[name = string("op_2852_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_2854_cast_fp16 = mul(x = var_2852_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2854_cast_fp16")];
            int32 var_2856 = const()[name = string("op_2856"), val = int32(-2)];
            bool var_2857_interleave_0 = const()[name = string("op_2857_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_2857_cast_fp16 = concat(axis = var_2856, interleave = var_2857_interleave_0, values = (var_2854_cast_fp16, var_2852_cast_fp16_0))[name = string("op_2857_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_2858_cast_fp16 = mul(x = var_2857_cast_fp16, y = var_459_cast_fp16)[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_75_cast_fp16 = add(x = var_2851_cast_fp16, y = var_2858_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_155")];
            tensor<fp16, [14, 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_96)[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_98_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_98 = read_state(input = key_cache)[name = string("coreml_update_state_98")];
            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_2834_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [14, 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_97)[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_99_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_99 = read_state(input = value_cache)[name = string("coreml_update_state_99")];
            tensor<int32, [4]> var_2928_begin_0 = const()[name = string("op_2928_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2928_end_0 = const()[name = string("op_2928_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2928_end_mask_0 = const()[name = string("op_2928_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2928_cast_fp16 = slice_by_index(begin = var_2928_begin_0, end = var_2928_end_0, end_mask = var_2928_end_mask_0, x = coreml_update_state_98)[name = string("op_2928_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2931_axis_0 = const()[name = string("op_2931_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_1 = split(axis = var_2931_axis_0, split_sizes = tile_14, x = var_2928_cast_fp16)[name = string("op_2931_cast_fp16")];
            tensor<int32, [4]> var_2938_begin_0 = const()[name = string("op_2938_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2938_end_0 = const()[name = string("op_2938_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2938_end_mask_0 = const()[name = string("op_2938_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2938_cast_fp16 = slice_by_index(begin = var_2938_begin_0, end = var_2938_end_0, end_mask = var_2938_end_mask_0, x = coreml_update_state_99)[name = string("op_2938_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2941_axis_0 = const()[name = string("op_2941_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_1 = split(axis = var_2941_axis_0, split_sizes = tile_15, x = var_2938_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<int32, [2]> var_2944_split_sizes_0 = const()[name = string("op_2944_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2944_axis_0 = const()[name = string("op_2944_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_2944_0, tensor<fp16, [1, 8, 128, 16]> var_2944_1 = split(axis = var_2944_axis_0, split_sizes = var_2944_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2944")];
            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_2931_cast_fp16_0, y = var_2944_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2947_to_fp16 = const()[name = string("op_2947_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_2947_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_2951 = const()[name = string("op_2951"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_119_cast_fp16 = softmax(axis = var_2951, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2957_transpose_x_1 = const()[name = string("op_2957_transpose_x_1"), val = bool(true)];
            bool var_2957_transpose_y_1 = const()[name = string("op_2957_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_2957_cast_fp16 = matmul(transpose_x = var_2957_transpose_x_1, transpose_y = var_2957_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2941_cast_fp16_0)[name = string("op_2957_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_2931_cast_fp16_1, y = var_2944_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2959_to_fp16 = const()[name = string("op_2959_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_2959_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_2963 = const()[name = string("op_2963"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_127_cast_fp16 = softmax(axis = var_2963, 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_2941_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2971 = const()[name = string("op_2971"), 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_2971, interleave = attn_output_59_interleave_0, values = (var_2957_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2975_perm_0 = const()[name = string("op_2975_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_2975_cast_fp16 = transpose(perm = var_2975_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2975_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_3008_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3008_cast_fp16")];
            int32 var_3006 = const()[name = string("op_3006"), 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_3006, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3008_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(802816128)))];
            fp16 var_3018_to_fp16 = const()[name = string("op_3018_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_3018_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3029_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3029_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_3029_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_3046_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3046_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(802824384)))];
            tensor<int32, [2]> var_3052_strides_0 = const()[name = string("op_3052_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3052_pad_type_0 = const()[name = string("op_3052_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3052_pad_0 = const()[name = string("op_3052_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3052_dilations_0 = const()[name = string("op_3052_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3052_groups_0 = const()[name = string("op_3052_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_3052_cast_fp16 = conv(dilations = var_3052_dilations_0, groups = var_3052_groups_0, pad = var_3052_pad_0, pad_type = var_3052_pad_type_0, strides = var_3052_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_3029_cast_fp16_0)[name = string("op_3052_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_79_cast_fp16 = mul(x = var_3046_cast_fp16, y = var_3052_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(827990272)))];
            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_3070_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3070_cast_fp16")];
            int32 var_3068 = const()[name = string("op_3068"), 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_3068, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3070_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(853156160)))];
            fp16 var_3080_to_fp16 = const()[name = string("op_3080_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_3080_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3091_split_sizes_0 = const()[name = string("op_3091_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3091_axis_0 = const()[name = string("op_3091_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3091_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3091_cast_fp16_1 = split(axis = var_3091_axis_0, split_sizes = var_3091_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3091_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(853164416)))];
            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_3091_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(861553088)))];
            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_3091_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            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_cast_fp16, x = var_3091_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_3148_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3148_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_3155_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3155_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3159_cast_fp16 = mul(x = x_81_cast_fp16, y = var_452_cast_fp16)[name = string("op_3159_cast_fp16")];
            tensor<int32, [2]> var_3160_split_sizes_0 = const()[name = string("op_3160_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3160_axis_0 = const()[name = string("op_3160_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_3160_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_3160_cast_fp16_1 = split(axis = var_3160_axis_0, split_sizes = var_3160_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3160_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_3162_cast_fp16 = mul(x = var_3160_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3162_cast_fp16")];
            int32 var_3164 = const()[name = string("op_3164"), val = int32(-2)];
            bool var_3165_interleave_0 = const()[name = string("op_3165_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_3165_cast_fp16 = concat(axis = var_3164, interleave = var_3165_interleave_0, values = (var_3162_cast_fp16, var_3160_cast_fp16_0))[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3166_cast_fp16 = mul(x = var_3165_cast_fp16, y = var_459_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_51_cast_fp16 = add(x = var_3159_cast_fp16, y = var_3166_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3172_cast_fp16 = mul(x = var_3148_cast_fp16, y = var_452_cast_fp16)[name = string("op_3172_cast_fp16")];
            tensor<int32, [2]> var_3173_split_sizes_0 = const()[name = string("op_3173_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3173_axis_0 = const()[name = string("op_3173_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_3173_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_3173_cast_fp16_1 = split(axis = var_3173_axis_0, split_sizes = var_3173_split_sizes_0, x = var_3148_cast_fp16)[name = string("op_3173_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_3175_cast_fp16 = mul(x = var_3173_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3175_cast_fp16")];
            int32 var_3177 = const()[name = string("op_3177"), val = int32(-2)];
            bool var_3178_interleave_0 = const()[name = string("op_3178_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_3178_cast_fp16 = concat(axis = var_3177, interleave = var_3178_interleave_0, values = (var_3175_cast_fp16, var_3173_cast_fp16_0))[name = string("op_3178_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3179_cast_fp16 = mul(x = var_3178_cast_fp16, y = var_459_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_85_cast_fp16 = add(x = var_3172_cast_fp16, y = var_3179_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_152")];
            tensor<fp16, [14, 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_98)[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_100_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_100 = read_state(input = key_cache)[name = string("coreml_update_state_100")];
            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_3155_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [14, 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_99)[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_101_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_101 = read_state(input = value_cache)[name = string("coreml_update_state_101")];
            tensor<int32, [4]> var_3249_begin_0 = const()[name = string("op_3249_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3249_end_0 = const()[name = string("op_3249_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3249_end_mask_0 = const()[name = string("op_3249_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3249_cast_fp16 = slice_by_index(begin = var_3249_begin_0, end = var_3249_end_0, end_mask = var_3249_end_mask_0, x = coreml_update_state_100)[name = string("op_3249_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3252_axis_0 = const()[name = string("op_3252_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_1 = split(axis = var_3252_axis_0, split_sizes = tile_16, x = var_3249_cast_fp16)[name = string("op_3252_cast_fp16")];
            tensor<int32, [4]> var_3259_begin_0 = const()[name = string("op_3259_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3259_end_0 = const()[name = string("op_3259_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3259_end_mask_0 = const()[name = string("op_3259_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3259_cast_fp16 = slice_by_index(begin = var_3259_begin_0, end = var_3259_end_0, end_mask = var_3259_end_mask_0, x = coreml_update_state_101)[name = string("op_3259_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3262_axis_0 = const()[name = string("op_3262_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_1 = split(axis = var_3262_axis_0, split_sizes = tile_17, x = var_3259_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<int32, [2]> var_3265_split_sizes_0 = const()[name = string("op_3265_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3265_axis_0 = const()[name = string("op_3265_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_3265_0, tensor<fp16, [1, 8, 128, 16]> var_3265_1 = split(axis = var_3265_axis_0, split_sizes = var_3265_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3265")];
            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_3252_cast_fp16_0, y = var_3265_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3268_to_fp16 = const()[name = string("op_3268_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_3268_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_3272 = const()[name = string("op_3272"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_135_cast_fp16 = softmax(axis = var_3272, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3278_transpose_x_1 = const()[name = string("op_3278_transpose_x_1"), val = bool(true)];
            bool var_3278_transpose_y_1 = const()[name = string("op_3278_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_3278_cast_fp16 = matmul(transpose_x = var_3278_transpose_x_1, transpose_y = var_3278_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3262_cast_fp16_0)[name = string("op_3278_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_3252_cast_fp16_1, y = var_3265_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3280_to_fp16 = const()[name = string("op_3280_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_3280_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_3284 = const()[name = string("op_3284"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_143_cast_fp16 = softmax(axis = var_3284, x = attn_weights_141_cast_fp16)[name = string("attn_weights_143_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_143_cast_fp16, y = var_3262_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3292 = const()[name = string("op_3292"), 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_3292, interleave = attn_output_67_interleave_0, values = (var_3278_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3296_perm_0 = const()[name = string("op_3296_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_3296_cast_fp16 = transpose(perm = var_3296_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3296_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            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_cast_fp16, x = attn_output_71_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_3329_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3329_cast_fp16")];
            int32 var_3327 = const()[name = string("op_3327"), 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_3327, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3329_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(862601728)))];
            fp16 var_3339_to_fp16 = const()[name = string("op_3339_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_3339_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3350_split_sizes_0 = const()[name = string("op_3350_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3350_axis_0 = const()[name = string("op_3350_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3350_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3350_cast_fp16_1 = split(axis = var_3350_axis_0, split_sizes = var_3350_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3350_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_3367_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3367_cast_fp16")];
            tensor<int32, [2]> var_3373_strides_0 = const()[name = string("op_3373_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3373_pad_type_0 = const()[name = string("op_3373_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3373_pad_0 = const()[name = string("op_3373_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3373_dilations_0 = const()[name = string("op_3373_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3373_groups_0 = const()[name = string("op_3373_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_3373_cast_fp16 = conv(dilations = var_3373_dilations_0, groups = var_3373_groups_0, pad = var_3373_pad_0, pad_type = var_3373_pad_type_0, strides = var_3373_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("op_3373_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_89_cast_fp16 = mul(x = var_3367_cast_fp16, y = var_3373_cast_fp16)[name = string("x_89_cast_fp16")];
            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_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_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_3391_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3391_cast_fp16")];
            int32 var_3389 = const()[name = string("op_3389"), 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_3389, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3391_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_37_gamma_0_to_fp16 = const()[name = string("out_37_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862609984)))];
            fp16 var_3401_to_fp16 = const()[name = string("op_3401_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_3401_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3412_split_sizes_0 = const()[name = string("op_3412_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3412_axis_0 = const()[name = string("op_3412_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3412_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3412_cast_fp16_1 = split(axis = var_3412_axis_0, split_sizes = var_3412_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3412_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862618240)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(871006912)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3412_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_3469_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3469_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_3476_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3480_cast_fp16 = mul(x = x_91_cast_fp16, y = var_452_cast_fp16)[name = string("op_3480_cast_fp16")];
            tensor<int32, [2]> var_3481_split_sizes_0 = const()[name = string("op_3481_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3481_axis_0 = const()[name = string("op_3481_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_3481_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_3481_cast_fp16_1 = split(axis = var_3481_axis_0, split_sizes = var_3481_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3481_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_3483_cast_fp16 = mul(x = var_3481_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3483_cast_fp16")];
            int32 var_3485 = const()[name = string("op_3485"), val = int32(-2)];
            bool var_3486_interleave_0 = const()[name = string("op_3486_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_3486_cast_fp16 = concat(axis = var_3485, interleave = var_3486_interleave_0, values = (var_3483_cast_fp16, var_3481_cast_fp16_0))[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3487_cast_fp16 = mul(x = var_3486_cast_fp16, y = var_459_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_57_cast_fp16 = add(x = var_3480_cast_fp16, y = var_3487_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3493_cast_fp16 = mul(x = var_3469_cast_fp16, y = var_452_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [2]> var_3494_split_sizes_0 = const()[name = string("op_3494_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3494_axis_0 = const()[name = string("op_3494_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_3494_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_3494_cast_fp16_1 = split(axis = var_3494_axis_0, split_sizes = var_3494_split_sizes_0, x = var_3469_cast_fp16)[name = string("op_3494_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_3496_cast_fp16 = mul(x = var_3494_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3496_cast_fp16")];
            int32 var_3498 = const()[name = string("op_3498"), val = int32(-2)];
            bool var_3499_interleave_0 = const()[name = string("op_3499_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_3499_cast_fp16 = concat(axis = var_3498, interleave = var_3499_interleave_0, values = (var_3496_cast_fp16, var_3494_cast_fp16_0))[name = string("op_3499_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3500_cast_fp16 = mul(x = var_3499_cast_fp16, y = var_459_cast_fp16)[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_95_cast_fp16 = add(x = var_3493_cast_fp16, y = var_3500_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_100)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_102_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_102 = read_state(input = key_cache)[name = string("coreml_update_state_102")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3476_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_101)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_103_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_103 = read_state(input = value_cache)[name = string("coreml_update_state_103")];
            tensor<int32, [4]> var_3570_begin_0 = const()[name = string("op_3570_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3570_end_0 = const()[name = string("op_3570_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3570_end_mask_0 = const()[name = string("op_3570_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = coreml_update_state_102)[name = string("op_3570_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3573_axis_0 = const()[name = string("op_3573_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_1 = split(axis = var_3573_axis_0, split_sizes = tile_18, x = var_3570_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [4]> var_3580_begin_0 = const()[name = string("op_3580_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3580_end_0 = const()[name = string("op_3580_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3580_end_mask_0 = const()[name = string("op_3580_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3580_cast_fp16 = slice_by_index(begin = var_3580_begin_0, end = var_3580_end_0, end_mask = var_3580_end_mask_0, x = coreml_update_state_103)[name = string("op_3580_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3583_axis_0 = const()[name = string("op_3583_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_1 = split(axis = var_3583_axis_0, split_sizes = tile_19, x = var_3580_cast_fp16)[name = string("op_3583_cast_fp16")];
            tensor<int32, [2]> var_3586_split_sizes_0 = const()[name = string("op_3586_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3586_axis_0 = const()[name = string("op_3586_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_3586_0, tensor<fp16, [1, 8, 128, 16]> var_3586_1 = split(axis = var_3586_axis_0, split_sizes = var_3586_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3586")];
            bool attn_weights_145_transpose_x_0 = const()[name = string("attn_weights_145_transpose_x_0"), val = bool(false)];
            bool attn_weights_145_transpose_y_0 = const()[name = string("attn_weights_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_145_cast_fp16 = matmul(transpose_x = attn_weights_145_transpose_x_0, transpose_y = attn_weights_145_transpose_y_0, x = var_3573_cast_fp16_0, y = var_3586_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3589_to_fp16 = const()[name = string("op_3589_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_147_cast_fp16 = mul(x = attn_weights_145_cast_fp16, y = var_3589_to_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_149_cast_fp16 = add(x = attn_weights_147_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_149_cast_fp16")];
            int32 var_3593 = const()[name = string("op_3593"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_151_cast_fp16 = softmax(axis = var_3593, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3599_transpose_x_1 = const()[name = string("op_3599_transpose_x_1"), val = bool(true)];
            bool var_3599_transpose_y_1 = const()[name = string("op_3599_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_3599_cast_fp16 = matmul(transpose_x = var_3599_transpose_x_1, transpose_y = var_3599_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3583_cast_fp16_0)[name = string("op_3599_cast_fp16")];
            bool attn_weights_153_transpose_x_0 = const()[name = string("attn_weights_153_transpose_x_0"), val = bool(false)];
            bool attn_weights_153_transpose_y_0 = const()[name = string("attn_weights_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_153_cast_fp16 = matmul(transpose_x = attn_weights_153_transpose_x_0, transpose_y = attn_weights_153_transpose_y_0, x = var_3573_cast_fp16_1, y = var_3586_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3601_to_fp16 = const()[name = string("op_3601_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_155_cast_fp16 = mul(x = attn_weights_153_cast_fp16, y = var_3601_to_fp16)[name = string("attn_weights_155_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_157_cast_fp16 = add(x = attn_weights_155_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_157_cast_fp16")];
            int32 var_3605 = const()[name = string("op_3605"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_159_cast_fp16 = softmax(axis = var_3605, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3583_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3613 = const()[name = string("op_3613"), val = int32(1)];
            bool attn_output_75_interleave_0 = const()[name = string("attn_output_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_75_cast_fp16 = concat(axis = var_3613, interleave = attn_output_75_interleave_0, values = (var_3599_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3617_perm_0 = const()[name = string("op_3617_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_3617_cast_fp16 = transpose(perm = var_3617_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3617_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_3650_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3650_cast_fp16")];
            int32 var_3648 = const()[name = string("op_3648"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_77_cast_fp16 = concat(axis = var_3648, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3650_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872055552)))];
            fp16 var_3660_to_fp16 = const()[name = string("op_3660_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3660_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3671_split_sizes_0 = const()[name = string("op_3671_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3671_axis_0 = const()[name = string("op_3671_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3671_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3671_cast_fp16_1 = split(axis = var_3671_axis_0, split_sizes = var_3671_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3671_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_3688_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3688_cast_fp16")];
            tensor<int32, [2]> var_3694_strides_0 = const()[name = string("op_3694_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3694_pad_type_0 = const()[name = string("op_3694_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3694_pad_0 = const()[name = string("op_3694_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3694_dilations_0 = const()[name = string("op_3694_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3694_groups_0 = const()[name = string("op_3694_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_3694_cast_fp16 = conv(dilations = var_3694_dilations_0, groups = var_3694_groups_0, pad = var_3694_pad_0, pad_type = var_3694_pad_type_0, strides = var_3694_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("op_3694_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_99_cast_fp16 = mul(x = var_3688_cast_fp16, y = var_3694_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_3712_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3712_cast_fp16")];
            int32 var_3710 = const()[name = string("op_3710"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_81_cast_fp16 = concat(axis = var_3710, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3712_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872063808)))];
            fp16 var_3722_to_fp16 = const()[name = string("op_3722_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3722_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3733_split_sizes_0 = const()[name = string("op_3733_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3733_axis_0 = const()[name = string("op_3733_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3733_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3733_cast_fp16_1 = split(axis = var_3733_axis_0, split_sizes = var_3733_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872072064)))];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(880460736)))];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_cast_fp16, x = var_3733_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_3790_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3790_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_3797_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3797_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3801_cast_fp16 = mul(x = x_101_cast_fp16, y = var_452_cast_fp16)[name = string("op_3801_cast_fp16")];
            tensor<int32, [2]> var_3802_split_sizes_0 = const()[name = string("op_3802_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3802_axis_0 = const()[name = string("op_3802_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_3802_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_3802_cast_fp16_1 = split(axis = var_3802_axis_0, split_sizes = var_3802_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3802_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_3804_cast_fp16 = mul(x = var_3802_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3804_cast_fp16")];
            int32 var_3806 = const()[name = string("op_3806"), val = int32(-2)];
            bool var_3807_interleave_0 = const()[name = string("op_3807_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_3807_cast_fp16 = concat(axis = var_3806, interleave = var_3807_interleave_0, values = (var_3804_cast_fp16, var_3802_cast_fp16_0))[name = string("op_3807_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_3808_cast_fp16 = mul(x = var_3807_cast_fp16, y = var_459_cast_fp16)[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_63_cast_fp16 = add(x = var_3801_cast_fp16, y = var_3808_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3814_cast_fp16 = mul(x = var_3790_cast_fp16, y = var_452_cast_fp16)[name = string("op_3814_cast_fp16")];
            tensor<int32, [2]> var_3815_split_sizes_0 = const()[name = string("op_3815_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3815_axis_0 = const()[name = string("op_3815_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_3815_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_3815_cast_fp16_1 = split(axis = var_3815_axis_0, split_sizes = var_3815_split_sizes_0, x = var_3790_cast_fp16)[name = string("op_3815_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_3817_cast_fp16 = mul(x = var_3815_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3817_cast_fp16")];
            int32 var_3819 = const()[name = string("op_3819"), val = int32(-2)];
            bool var_3820_interleave_0 = const()[name = string("op_3820_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_3820_cast_fp16 = concat(axis = var_3819, interleave = var_3820_interleave_0, values = (var_3817_cast_fp16, var_3815_cast_fp16_0))[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_3821_cast_fp16 = mul(x = var_3820_cast_fp16, y = var_459_cast_fp16)[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_105_cast_fp16 = add(x = var_3814_cast_fp16, y = var_3821_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_102)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_104_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_104 = read_state(input = key_cache)[name = string("coreml_update_state_104")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3797_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_103)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_105_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_105 = read_state(input = value_cache)[name = string("coreml_update_state_105")];
            tensor<int32, [4]> var_3891_begin_0 = const()[name = string("op_3891_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3891_end_0 = const()[name = string("op_3891_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3891_end_mask_0 = const()[name = string("op_3891_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3891_cast_fp16 = slice_by_index(begin = var_3891_begin_0, end = var_3891_end_0, end_mask = var_3891_end_mask_0, x = coreml_update_state_104)[name = string("op_3891_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3894_axis_0 = const()[name = string("op_3894_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_1 = split(axis = var_3894_axis_0, split_sizes = tile_20, x = var_3891_cast_fp16)[name = string("op_3894_cast_fp16")];
            tensor<int32, [4]> var_3901_begin_0 = const()[name = string("op_3901_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3901_end_0 = const()[name = string("op_3901_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3901_end_mask_0 = const()[name = string("op_3901_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3901_cast_fp16 = slice_by_index(begin = var_3901_begin_0, end = var_3901_end_0, end_mask = var_3901_end_mask_0, x = coreml_update_state_105)[name = string("op_3901_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3904_axis_0 = const()[name = string("op_3904_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_1 = split(axis = var_3904_axis_0, split_sizes = tile_21, x = var_3901_cast_fp16)[name = string("op_3904_cast_fp16")];
            tensor<int32, [2]> var_3907_split_sizes_0 = const()[name = string("op_3907_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3907_axis_0 = const()[name = string("op_3907_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_3907_0, tensor<fp16, [1, 8, 128, 16]> var_3907_1 = split(axis = var_3907_axis_0, split_sizes = var_3907_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3907")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3894_cast_fp16_0, y = var_3907_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3910_to_fp16 = const()[name = string("op_3910_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3910_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3914 = const()[name = string("op_3914"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_167_cast_fp16 = softmax(axis = var_3914, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3920_transpose_x_1 = const()[name = string("op_3920_transpose_x_1"), val = bool(true)];
            bool var_3920_transpose_y_1 = const()[name = string("op_3920_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_3920_cast_fp16 = matmul(transpose_x = var_3920_transpose_x_1, transpose_y = var_3920_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3904_cast_fp16_0)[name = string("op_3920_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3894_cast_fp16_1, y = var_3907_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3922_to_fp16 = const()[name = string("op_3922_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3922_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3926 = const()[name = string("op_3926"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_175_cast_fp16 = softmax(axis = var_3926, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3904_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3934 = const()[name = string("op_3934"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_83_cast_fp16 = concat(axis = var_3934, interleave = attn_output_83_interleave_0, values = (var_3920_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3938_perm_0 = const()[name = string("op_3938_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_3938_cast_fp16 = transpose(perm = var_3938_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3938_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_3971_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3971_cast_fp16")];
            int32 var_3969 = const()[name = string("op_3969"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_85_cast_fp16 = concat(axis = var_3969, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3971_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881509376)))];
            fp16 var_3981_to_fp16 = const()[name = string("op_3981_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3981_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3992_split_sizes_0 = const()[name = string("op_3992_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3992_axis_0 = const()[name = string("op_3992_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_3992_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_3992_cast_fp16_1 = split(axis = var_3992_axis_0, split_sizes = var_3992_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3992_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_4009_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [2]> var_4015_strides_0 = const()[name = string("op_4015_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4015_pad_type_0 = const()[name = string("op_4015_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4015_pad_0 = const()[name = string("op_4015_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4015_dilations_0 = const()[name = string("op_4015_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4015_groups_0 = const()[name = string("op_4015_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_4015_cast_fp16 = conv(dilations = var_4015_dilations_0, groups = var_4015_groups_0, pad = var_4015_pad_0, pad_type = var_4015_pad_type_0, strides = var_4015_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("op_4015_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_109_cast_fp16 = mul(x = var_4009_cast_fp16, y = var_4015_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_4033_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4033_cast_fp16")];
            int32 var_4031 = const()[name = string("op_4031"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_89_cast_fp16 = concat(axis = var_4031, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4033_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881517632)))];
            fp16 var_4043_to_fp16 = const()[name = string("op_4043_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4043_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4054_split_sizes_0 = const()[name = string("op_4054_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4054_axis_0 = const()[name = string("op_4054_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_4054_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_4054_cast_fp16_1 = split(axis = var_4054_axis_0, split_sizes = var_4054_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_4111_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4111_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_4118_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4118_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_4122_cast_fp16 = mul(x = x_111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4122_cast_fp16")];
            tensor<int32, [2]> var_4123_split_sizes_0 = const()[name = string("op_4123_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4123_axis_0 = const()[name = string("op_4123_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_4123_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_4123_cast_fp16_1 = split(axis = var_4123_axis_0, split_sizes = var_4123_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4123_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_4125_cast_fp16 = mul(x = var_4123_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4125_cast_fp16")];
            int32 var_4127 = const()[name = string("op_4127"), val = int32(-2)];
            bool var_4128_interleave_0 = const()[name = string("op_4128_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_4128_cast_fp16 = concat(axis = var_4127, interleave = var_4128_interleave_0, values = (var_4125_cast_fp16, var_4123_cast_fp16_0))[name = string("op_4128_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_4129_cast_fp16 = mul(x = var_4128_cast_fp16, y = var_459_cast_fp16)[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_69_cast_fp16 = add(x = var_4122_cast_fp16, y = var_4129_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_4135_cast_fp16 = mul(x = var_4111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4135_cast_fp16")];
            tensor<int32, [2]> var_4136_split_sizes_0 = const()[name = string("op_4136_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4136_axis_0 = const()[name = string("op_4136_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_4136_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_4136_cast_fp16_1 = split(axis = var_4136_axis_0, split_sizes = var_4136_split_sizes_0, x = var_4111_cast_fp16)[name = string("op_4136_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_4138_cast_fp16 = mul(x = var_4136_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4138_cast_fp16")];
            int32 var_4140 = const()[name = string("op_4140"), val = int32(-2)];
            bool var_4141_interleave_0 = const()[name = string("op_4141_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_4141_cast_fp16 = concat(axis = var_4140, interleave = var_4141_interleave_0, values = (var_4138_cast_fp16, var_4136_cast_fp16_0))[name = string("op_4141_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_4142_cast_fp16 = mul(x = var_4141_cast_fp16, y = var_459_cast_fp16)[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_115_cast_fp16 = add(x = var_4135_cast_fp16, y = var_4142_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_104)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_106_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_106 = read_state(input = key_cache)[name = string("coreml_update_state_106")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4118_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_105)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_107_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_107 = read_state(input = value_cache)[name = string("coreml_update_state_107")];
            tensor<int32, [4]> var_4212_begin_0 = const()[name = string("op_4212_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4212_end_0 = const()[name = string("op_4212_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4212_end_mask_0 = const()[name = string("op_4212_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4212_cast_fp16 = slice_by_index(begin = var_4212_begin_0, end = var_4212_end_0, end_mask = var_4212_end_mask_0, x = coreml_update_state_106)[name = string("op_4212_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4215_axis_0 = const()[name = string("op_4215_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_1 = split(axis = var_4215_axis_0, split_sizes = tile_22, x = var_4212_cast_fp16)[name = string("op_4215_cast_fp16")];
            tensor<int32, [4]> var_4222_begin_0 = const()[name = string("op_4222_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4222_end_0 = const()[name = string("op_4222_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4222_end_mask_0 = const()[name = string("op_4222_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4222_cast_fp16 = slice_by_index(begin = var_4222_begin_0, end = var_4222_end_0, end_mask = var_4222_end_mask_0, x = coreml_update_state_107)[name = string("op_4222_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4225_axis_0 = const()[name = string("op_4225_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_1 = split(axis = var_4225_axis_0, split_sizes = tile_23, x = var_4222_cast_fp16)[name = string("op_4225_cast_fp16")];
            tensor<int32, [2]> var_4228_split_sizes_0 = const()[name = string("op_4228_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4228_axis_0 = const()[name = string("op_4228_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_4228_0, tensor<fp16, [1, 8, 128, 16]> var_4228_1 = split(axis = var_4228_axis_0, split_sizes = var_4228_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4228")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4215_cast_fp16_0, y = var_4228_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4231_to_fp16 = const()[name = string("op_4231_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4231_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4235 = const()[name = string("op_4235"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_183_cast_fp16 = softmax(axis = var_4235, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4241_transpose_x_1 = const()[name = string("op_4241_transpose_x_1"), val = bool(true)];
            bool var_4241_transpose_y_1 = const()[name = string("op_4241_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_4241_cast_fp16 = matmul(transpose_x = var_4241_transpose_x_1, transpose_y = var_4241_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4225_cast_fp16_0)[name = string("op_4241_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4215_cast_fp16_1, y = var_4228_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4243_to_fp16 = const()[name = string("op_4243_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4243_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4247 = const()[name = string("op_4247"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_191_cast_fp16 = softmax(axis = var_4247, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4225_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4255 = const()[name = string("op_4255"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_91_cast_fp16 = concat(axis = var_4255, interleave = attn_output_91_interleave_0, values = (var_4241_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4259_perm_0 = const()[name = string("op_4259_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_4259_cast_fp16 = transpose(perm = var_4259_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4259_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_4292_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4292_cast_fp16")];
            int32 var_4290 = const()[name = string("op_4290"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_93_cast_fp16 = concat(axis = var_4290, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4292_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881525888)))];
            fp16 var_4302_to_fp16 = const()[name = string("op_4302_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4302_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4313_split_sizes_0 = const()[name = string("op_4313_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4313_axis_0 = const()[name = string("op_4313_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_4313_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_4313_cast_fp16_1 = split(axis = var_4313_axis_0, split_sizes = var_4313_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4313_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_4330_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4330_cast_fp16")];
            tensor<int32, [2]> var_4336_strides_0 = const()[name = string("op_4336_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4336_pad_type_0 = const()[name = string("op_4336_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4336_pad_0 = const()[name = string("op_4336_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4336_dilations_0 = const()[name = string("op_4336_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4336_groups_0 = const()[name = string("op_4336_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_4336_cast_fp16 = conv(dilations = var_4336_dilations_0, groups = var_4336_groups_0, pad = var_4336_pad_0, pad_type = var_4336_pad_type_0, strides = var_4336_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("op_4336_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_119_cast_fp16 = mul(x = var_4330_cast_fp16, y = var_4336_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_cast_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_4354_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4354_cast_fp16")];
            int32 var_4352 = const()[name = string("op_4352"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_97_cast_fp16 = concat(axis = var_4352, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4354_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881534144)))];
            fp16 var_4364_to_fp16 = const()[name = string("op_4364_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4364_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4375_split_sizes_0 = const()[name = string("op_4375_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4375_axis_0 = const()[name = string("op_4375_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_4375_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_4375_cast_fp16_1 = split(axis = var_4375_axis_0, split_sizes = var_4375_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4375_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_4432_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4432_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_4439_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4439_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_4443_cast_fp16 = mul(x = x_121_cast_fp16, y = var_452_cast_fp16)[name = string("op_4443_cast_fp16")];
            tensor<int32, [2]> var_4444_split_sizes_0 = const()[name = string("op_4444_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4444_axis_0 = const()[name = string("op_4444_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_4444_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_4444_cast_fp16_1 = split(axis = var_4444_axis_0, split_sizes = var_4444_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4444_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_4446_cast_fp16 = mul(x = var_4444_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4446_cast_fp16")];
            int32 var_4448 = const()[name = string("op_4448"), val = int32(-2)];
            bool var_4449_interleave_0 = const()[name = string("op_4449_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_4449_cast_fp16 = concat(axis = var_4448, interleave = var_4449_interleave_0, values = (var_4446_cast_fp16, var_4444_cast_fp16_0))[name = string("op_4449_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_4450_cast_fp16 = mul(x = var_4449_cast_fp16, y = var_459_cast_fp16)[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_75_cast_fp16 = add(x = var_4443_cast_fp16, y = var_4450_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_4456_cast_fp16 = mul(x = var_4432_cast_fp16, y = var_452_cast_fp16)[name = string("op_4456_cast_fp16")];
            tensor<int32, [2]> var_4457_split_sizes_0 = const()[name = string("op_4457_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4457_axis_0 = const()[name = string("op_4457_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_4457_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_4457_cast_fp16_1 = split(axis = var_4457_axis_0, split_sizes = var_4457_split_sizes_0, x = var_4432_cast_fp16)[name = string("op_4457_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_4459_cast_fp16 = mul(x = var_4457_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4459_cast_fp16")];
            int32 var_4461 = const()[name = string("op_4461"), val = int32(-2)];
            bool var_4462_interleave_0 = const()[name = string("op_4462_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_4462_cast_fp16 = concat(axis = var_4461, interleave = var_4462_interleave_0, values = (var_4459_cast_fp16, var_4457_cast_fp16_0))[name = string("op_4462_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_4463_cast_fp16 = mul(x = var_4462_cast_fp16, y = var_459_cast_fp16)[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_125_cast_fp16 = add(x = var_4456_cast_fp16, y = var_4463_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_106)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_108_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_108 = read_state(input = key_cache)[name = string("coreml_update_state_108")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4439_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_107)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_109_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_109 = read_state(input = value_cache)[name = string("coreml_update_state_109")];
            tensor<int32, [4]> var_4533_begin_0 = const()[name = string("op_4533_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4533_end_0 = const()[name = string("op_4533_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4533_end_mask_0 = const()[name = string("op_4533_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4533_cast_fp16 = slice_by_index(begin = var_4533_begin_0, end = var_4533_end_0, end_mask = var_4533_end_mask_0, x = coreml_update_state_108)[name = string("op_4533_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4536_axis_0 = const()[name = string("op_4536_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_1 = split(axis = var_4536_axis_0, split_sizes = tile_24, x = var_4533_cast_fp16)[name = string("op_4536_cast_fp16")];
            tensor<int32, [4]> var_4543_begin_0 = const()[name = string("op_4543_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4543_end_0 = const()[name = string("op_4543_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4543_end_mask_0 = const()[name = string("op_4543_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4543_cast_fp16 = slice_by_index(begin = var_4543_begin_0, end = var_4543_end_0, end_mask = var_4543_end_mask_0, x = coreml_update_state_109)[name = string("op_4543_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4546_axis_0 = const()[name = string("op_4546_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_1 = split(axis = var_4546_axis_0, split_sizes = tile_25, x = var_4543_cast_fp16)[name = string("op_4546_cast_fp16")];
            tensor<int32, [2]> var_4549_split_sizes_0 = const()[name = string("op_4549_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4549_axis_0 = const()[name = string("op_4549_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_4549_0, tensor<fp16, [1, 8, 128, 16]> var_4549_1 = split(axis = var_4549_axis_0, split_sizes = var_4549_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4549")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4536_cast_fp16_0, y = var_4549_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4552_to_fp16 = const()[name = string("op_4552_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4552_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4556 = const()[name = string("op_4556"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_199_cast_fp16 = softmax(axis = var_4556, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4562_transpose_x_1 = const()[name = string("op_4562_transpose_x_1"), val = bool(true)];
            bool var_4562_transpose_y_1 = const()[name = string("op_4562_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_4562_cast_fp16 = matmul(transpose_x = var_4562_transpose_x_1, transpose_y = var_4562_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4546_cast_fp16_0)[name = string("op_4562_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4536_cast_fp16_1, y = var_4549_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4564_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4568 = const()[name = string("op_4568"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_207_cast_fp16 = softmax(axis = var_4568, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4546_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4576 = const()[name = string("op_4576"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_99_cast_fp16 = concat(axis = var_4576, interleave = attn_output_99_interleave_0, values = (var_4562_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4580_perm_0 = const()[name = string("op_4580_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_4580_cast_fp16 = transpose(perm = var_4580_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4580_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_4613_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4613_cast_fp16")];
            int32 var_4611 = const()[name = string("op_4611"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_101_cast_fp16 = concat(axis = var_4611, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4613_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881542400)))];
            fp16 var_4623_to_fp16 = const()[name = string("op_4623_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4623_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4634_split_sizes_0 = const()[name = string("op_4634_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4634_axis_0 = const()[name = string("op_4634_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_4634_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_4634_cast_fp16_1 = split(axis = var_4634_axis_0, split_sizes = var_4634_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4634_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_4651_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4651_cast_fp16")];
            tensor<int32, [2]> var_4657_strides_0 = const()[name = string("op_4657_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4657_pad_type_0 = const()[name = string("op_4657_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4657_pad_0 = const()[name = string("op_4657_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4657_dilations_0 = const()[name = string("op_4657_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4657_groups_0 = const()[name = string("op_4657_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_4657_cast_fp16 = conv(dilations = var_4657_dilations_0, groups = var_4657_groups_0, pad = var_4657_pad_0, pad_type = var_4657_pad_type_0, strides = var_4657_strides_0, weight = layers_12_mlp_up_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("op_4657_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_129_cast_fp16 = mul(x = var_4651_cast_fp16, y = var_4657_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_cast_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_4675_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4675_cast_fp16")];
            int32 var_4673 = const()[name = string("op_4673"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_105_cast_fp16 = concat(axis = var_4673, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4675_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881550656)))];
            fp16 var_4685_to_fp16 = const()[name = string("op_4685_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4685_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4696_split_sizes_0 = const()[name = string("op_4696_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4696_axis_0 = const()[name = string("op_4696_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_4696_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_4696_cast_fp16_1 = split(axis = var_4696_axis_0, split_sizes = var_4696_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4696_cast_fp16")];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 16]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 16]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_4753_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4753_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 16]> var_4760_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4760_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_4764_cast_fp16 = mul(x = x_131_cast_fp16, y = var_452_cast_fp16)[name = string("op_4764_cast_fp16")];
            tensor<int32, [2]> var_4765_split_sizes_0 = const()[name = string("op_4765_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4765_axis_0 = const()[name = string("op_4765_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 16]> var_4765_cast_fp16_0, tensor<fp16, [1, 16, 64, 16]> var_4765_cast_fp16_1 = split(axis = var_4765_axis_0, split_sizes = var_4765_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4765_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 16]> var_4767_cast_fp16 = mul(x = var_4765_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4767_cast_fp16")];
            int32 var_4769 = const()[name = string("op_4769"), val = int32(-2)];
            bool var_4770_interleave_0 = const()[name = string("op_4770_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 16]> var_4770_cast_fp16 = concat(axis = var_4769, interleave = var_4770_interleave_0, values = (var_4767_cast_fp16, var_4765_cast_fp16_0))[name = string("op_4770_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> var_4771_cast_fp16 = mul(x = var_4770_cast_fp16, y = var_459_cast_fp16)[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 16]> query_states_81_cast_fp16 = add(x = var_4764_cast_fp16, y = var_4771_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_4777_cast_fp16 = mul(x = var_4753_cast_fp16, y = var_452_cast_fp16)[name = string("op_4777_cast_fp16")];
            tensor<int32, [2]> var_4778_split_sizes_0 = const()[name = string("op_4778_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4778_axis_0 = const()[name = string("op_4778_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 16]> var_4778_cast_fp16_0, tensor<fp16, [1, 2, 64, 16]> var_4778_cast_fp16_1 = split(axis = var_4778_axis_0, split_sizes = var_4778_split_sizes_0, x = var_4753_cast_fp16)[name = string("op_4778_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 16]> var_4780_cast_fp16 = mul(x = var_4778_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4780_cast_fp16")];
            int32 var_4782 = const()[name = string("op_4782"), val = int32(-2)];
            bool var_4783_interleave_0 = const()[name = string("op_4783_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 16]> var_4783_cast_fp16 = concat(axis = var_4782, interleave = var_4783_interleave_0, values = (var_4780_cast_fp16, var_4778_cast_fp16_0))[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> var_4784_cast_fp16 = mul(x = var_4783_cast_fp16, y = var_459_cast_fp16)[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 16]> key_states_135_cast_fp16 = add(x = var_4777_cast_fp16, y = var_4784_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_108)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_110_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_110 = read_state(input = key_cache)[name = string("coreml_update_state_110")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 16, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4760_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_109)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_111_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_111 = read_state(input = value_cache)[name = string("coreml_update_state_111")];
            tensor<int32, [4]> var_4854_begin_0 = const()[name = string("op_4854_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4854_end_0 = const()[name = string("op_4854_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4854_end_mask_0 = const()[name = string("op_4854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4854_cast_fp16 = slice_by_index(begin = var_4854_begin_0, end = var_4854_end_0, end_mask = var_4854_end_mask_0, x = coreml_update_state_110)[name = string("op_4854_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4857_axis_0 = const()[name = string("op_4857_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_1 = split(axis = var_4857_axis_0, split_sizes = tile_26, x = var_4854_cast_fp16)[name = string("op_4857_cast_fp16")];
            tensor<int32, [4]> var_4864_begin_0 = const()[name = string("op_4864_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4864_end_0 = const()[name = string("op_4864_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4864_end_mask_0 = const()[name = string("op_4864_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4864_cast_fp16 = slice_by_index(begin = var_4864_begin_0, end = var_4864_end_0, end_mask = var_4864_end_mask_0, x = coreml_update_state_111)[name = string("op_4864_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4867_axis_0 = const()[name = string("op_4867_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_1 = split(axis = var_4867_axis_0, split_sizes = tile_27, x = var_4864_cast_fp16)[name = string("op_4867_cast_fp16")];
            tensor<int32, [2]> var_4870_split_sizes_0 = const()[name = string("op_4870_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4870_axis_0 = const()[name = string("op_4870_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 16]> var_4870_0, tensor<fp16, [1, 8, 128, 16]> var_4870_1 = split(axis = var_4870_axis_0, split_sizes = var_4870_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4870")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4857_cast_fp16_0, y = var_4870_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4873_to_fp16 = const()[name = string("op_4873_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4873_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4877 = const()[name = string("op_4877"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_215_cast_fp16 = softmax(axis = var_4877, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4883_transpose_x_1 = const()[name = string("op_4883_transpose_x_1"), val = bool(true)];
            bool var_4883_transpose_y_1 = const()[name = string("op_4883_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> var_4883_cast_fp16 = matmul(transpose_x = var_4883_transpose_x_1, transpose_y = var_4883_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4867_cast_fp16_0)[name = string("op_4883_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4857_cast_fp16_1, y = var_4870_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4885_to_fp16 = const()[name = string("op_4885_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4885_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4889 = const()[name = string("op_4889"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 16]> attn_weights_cast_fp16 = softmax(axis = var_4889, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4867_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4897 = const()[name = string("op_4897"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> attn_output_107_cast_fp16 = concat(axis = var_4897, interleave = attn_output_107_interleave_0, values = (var_4883_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4901_perm_0 = const()[name = string("op_4901_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 16]> var_4901_cast_fp16 = transpose(perm = var_4901_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 2048, 1, 16]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4901_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 16]> var_4934_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4934_cast_fp16")];
            int32 var_4932 = const()[name = string("op_4932"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 16]> doubled_109_cast_fp16 = concat(axis = var_4932, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4934_cast_fp16))[name = string("doubled_109_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(881558912)))];
            fp16 var_4944_to_fp16 = const()[name = string("op_4944_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 16]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_4944_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_4955_split_sizes_0 = const()[name = string("op_4955_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4955_axis_0 = const()[name = string("op_4955_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> var_4955_cast_fp16_0, tensor<fp16, [1, 2048, 1, 16]> var_4955_cast_fp16_1 = split(axis = var_4955_axis_0, split_sizes = var_4955_split_sizes_0, x = out_cast_fp16)[name = string("op_4955_cast_fp16")];
            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_13_mlp_gate_proj_weight_cast_fp16, x = var_4955_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> var_4972_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4972_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881567168)))];
            tensor<int32, [2]> var_4978_strides_0 = const()[name = string("op_4978_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4978_pad_type_0 = const()[name = string("op_4978_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4978_pad_0 = const()[name = string("op_4978_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4978_dilations_0 = const()[name = string("op_4978_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4978_groups_0 = const()[name = string("op_4978_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 16]> var_4978_cast_fp16 = conv(dilations = var_4978_dilations_0, groups = var_4978_groups_0, pad = var_4978_pad_0, pad_type = var_4978_pad_type_0, strides = var_4978_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4955_cast_fp16_0)[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 16]> x_cast_fp16 = mul(x = var_4972_cast_fp16, y = var_4978_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_13_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 16]> hidden_states = add(x = hidden_states_135_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_4987_cast_fp16")];
        } -> (hidden_states);
    func length_32<ios18>(tensor<fp32, [1, 2048, 1, 32]> inputs_embeds, state<tensor<fp16, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [32]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_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(64))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524992))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524416))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_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(525312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13120640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13108288))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_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(13126848))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651776))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651200))))[name = string("layers_2_self_attn_v_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(13652096))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26247424))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26235072))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_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(26253632))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26778560))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26777984))))[name = string("layers_3_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_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(26778880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30977408))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30973248))))[name = string("layers_3_self_attn_o_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(30979520))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43566656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43562496))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_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(43568768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093120))))[name = string("layers_4_self_attn_v_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(44094016))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48292544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48288384))))[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(48294656))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60889984))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60877632))))[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(60896192))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73491520))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73479168))))[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(73497728))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86084864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86080704))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_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(86086976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611328))))[name = string("layers_5_self_attn_v_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(86612224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90810752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90806592))))[name = string("layers_5_self_attn_o_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(90812864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103395840))))[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(103414400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997376))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_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(116003648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528000))))[name = string("layers_6_self_attn_v_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(116528896))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120727424))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120723264))))[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(120729536))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133324864))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133312512))))[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(133331072))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145926400))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145914048))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_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(145932608))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158519744))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158515584))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_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(158521856))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046784))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046208))))[name = string("layers_7_self_attn_v_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(159047104))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163245632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241472))))[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(163247744))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175843072))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175830720))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_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(175849280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176374208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176373632))))[name = string("layers_8_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_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(176374528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180573056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180568896))))[name = string("layers_8_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(180575168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193170496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193158144))))[name = string("layers_8_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(193176704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205772032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205759680))))[name = string("layers_8_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_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(205778240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218365376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218361216))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_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(218367488))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218892416))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218891840))))[name = string("layers_9_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_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(218892736))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223091264))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223087104))))[name = string("layers_9_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(223093376))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235688704))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235676352))))[name = string("layers_9_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(235694912))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248290240))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248277888))))[name = string("layers_9_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_9_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(248296448))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260883584))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260879424))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_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(260885696))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410624))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410048))))[name = string("layers_10_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_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(261410944))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265609472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265605312))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(265611584))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278206912))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278194560))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(278213120))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290808448))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290796096))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_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(290814656))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303401792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303397632))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(303403904))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307602432))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307598272))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_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(307604544))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308129472))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308128896))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_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(308129792))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654720))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654144))))[name = string("layers_11_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(308655040))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312853568))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312849408))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(312855680))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325451008))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325438656))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(325457216))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338052544))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040192))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_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(338058752))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350645888))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350641728))))[name = string("layers_11_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(350648000))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354846528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354842368))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_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(354848640))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355373568))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355372992))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_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(355373888))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898816))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898240))))[name = string("layers_12_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(355899136))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360097664))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360093504))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(360099776))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372695104))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372682752))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(372701312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385296640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385284288))))[name = string("layers_12_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_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(385302848))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397889984))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397885824))))[name = string("layers_12_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(397892096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402090624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402086464))))[name = string("layers_13_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_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(402092736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617088))))[name = string("layers_13_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_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(402617984))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142912))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142336))))[name = string("layers_13_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(403143232))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407341760))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407337600))))[name = string("layers_13_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_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(407343872))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419939200))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419926848))))[name = string("layers_13_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_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(419945408))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432532544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432528384))))[name = string("layers_13_mlp_down_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_424 = const()[name = string("op_424"), val = int32(0)];
            bool var_426_exclusive_0 = const()[name = string("op_426_exclusive_0"), val = bool(false)];
            bool var_426_reverse_0 = const()[name = string("op_426_reverse_0"), val = bool(false)];
            tensor<fp16, [32]> var_426_cast_fp16 = cumsum(axis = var_424, exclusive = var_426_exclusive_0, reverse = var_426_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_426_cast_fp16")];
            fp16 var_428_promoted_to_fp16 = const()[name = string("op_428_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [32]> position_offsets_cast_fp16 = sub(x = var_426_cast_fp16, y = var_428_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_431_axes_0 = const()[name = string("op_431_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 32]> var_431_cast_fp16 = expand_dims(axes = var_431_axes_0, x = position_offsets_cast_fp16)[name = string("op_431_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_431_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(432534656)))];
            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(440923328)))];
            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_450_perm_0 = const()[name = string("op_450_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_452_axes_0 = const()[name = string("op_452_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 32]> var_450_cast_fp16 = transpose(perm = var_450_perm_0, x = cos_1_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [1, 1, 128, 32]> var_452_cast_fp16 = expand_dims(axes = var_452_axes_0, x = var_450_cast_fp16)[name = string("op_452_cast_fp16")];
            tensor<int32, [3]> var_457_perm_0 = const()[name = string("op_457_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_459_axes_0 = const()[name = string("op_459_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 32]> var_457_cast_fp16 = transpose(perm = var_457_perm_0, x = sin_1_cast_fp16)[name = string("transpose_223")];
            tensor<fp16, [1, 1, 128, 32]> var_459_cast_fp16 = expand_dims(axes = var_459_axes_0, x = var_457_cast_fp16)[name = string("op_459_cast_fp16")];
            tensor<int32, [1]> var_478_axes_0 = const()[name = string("op_478_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 32, 1]> var_478 = expand_dims(axes = var_478_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_478")];
            tensor<int32, [1, 1, 2048]> var_471 = const()[name = string("op_471"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449312000)))];
            tensor<bool, [1, 32, 2048]> var_479 = greater(x = var_471, y = var_478)[name = string("op_479")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_486_axes_0 = const()[name = string("op_486_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 32, 2048]> var_479_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_479)[name = string("cast_17")];
            tensor<fp16, [1, 1, 32, 2048]> var_486_cast_fp16 = expand_dims(axes = var_486_axes_0, x = var_479_to_fp16)[name = string("op_486_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_490_promoted_to_fp16 = const()[name = string("op_490_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 32]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_486_cast_fp16)[name = string("transpose_222")];
            tensor<bool, [1, 1, 2048, 32]> var_491_cast_fp16 = equal(x = mask_cast_fp16, y = var_490_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            fp16 var_492_to_fp16 = const()[name = string("op_492_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 32]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_492_to_fp16, cond = var_491_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_502_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_502_cast_fp16")];
            int32 var_500 = const()[name = string("op_500"), 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_500, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_502_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(449320256)))];
            fp16 var_512_to_fp16 = const()[name = string("op_512_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_512_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_523_split_sizes_0 = const()[name = string("op_523_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_523_axis_0 = const()[name = string("op_523_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_523_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_523_cast_fp16_1 = split(axis = var_523_axis_0, split_sizes = var_523_split_sizes_0, x = out_1_cast_fp16)[name = string("op_523_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449328512)))];
            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_to_fp16, x = var_523_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457717184)))];
            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_to_fp16, x = var_523_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(458765824)))];
            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_523_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_580_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_580_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_587_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_587_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_591_cast_fp16 = mul(x = x_1_cast_fp16, y = var_452_cast_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [2]> var_592_split_sizes_0 = const()[name = string("op_592_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_592_axis_0 = const()[name = string("op_592_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_592_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_592_cast_fp16_1 = split(axis = var_592_axis_0, split_sizes = var_592_split_sizes_0, x = x_1_cast_fp16)[name = string("op_592_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_594_cast_fp16 = mul(x = var_592_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_594_cast_fp16")];
            int32 var_596 = const()[name = string("op_596"), val = int32(-2)];
            bool var_597_interleave_0 = const()[name = string("op_597_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_597_cast_fp16 = concat(axis = var_596, interleave = var_597_interleave_0, values = (var_594_cast_fp16, var_592_cast_fp16_0))[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_598_cast_fp16 = mul(x = var_597_cast_fp16, y = var_459_cast_fp16)[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_3_cast_fp16 = add(x = var_591_cast_fp16, y = var_598_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_604_cast_fp16 = mul(x = var_580_cast_fp16, y = var_452_cast_fp16)[name = string("op_604_cast_fp16")];
            tensor<int32, [2]> var_605_split_sizes_0 = const()[name = string("op_605_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_605_axis_0 = const()[name = string("op_605_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_605_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_605_cast_fp16_1 = split(axis = var_605_axis_0, split_sizes = var_605_split_sizes_0, x = var_580_cast_fp16)[name = string("op_605_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_607_cast_fp16 = mul(x = var_605_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_607_cast_fp16")];
            int32 var_609 = const()[name = string("op_609"), val = int32(-2)];
            bool var_610_interleave_0 = const()[name = string("op_610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_610_cast_fp16 = concat(axis = var_609, interleave = var_610_interleave_0, values = (var_607_cast_fp16, var_605_cast_fp16_0))[name = string("op_610_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_611_cast_fp16 = mul(x = var_610_cast_fp16, y = var_459_cast_fp16)[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_5_cast_fp16 = add(x = var_604_cast_fp16, y = var_611_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 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_221")];
            tensor<fp16, [14, 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_112_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_112 = read_state(input = key_cache)[name = string("coreml_update_state_112")];
            tensor<fp16, [14, 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_587_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [14, 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_113_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_113 = read_state(input = value_cache)[name = string("coreml_update_state_113")];
            tensor<int32, [4]> var_681_begin_0 = const()[name = string("op_681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_681_end_0 = const()[name = string("op_681_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_681_end_mask_0 = const()[name = string("op_681_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_681_cast_fp16 = slice_by_index(begin = var_681_begin_0, end = var_681_end_0, end_mask = var_681_end_mask_0, x = coreml_update_state_112)[name = string("op_681_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_684_axis_0 = const()[name = string("op_684_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_1 = split(axis = var_684_axis_0, split_sizes = tile_0, x = var_681_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [4]> var_691_begin_0 = const()[name = string("op_691_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_691_end_0 = const()[name = string("op_691_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_691_end_mask_0 = const()[name = string("op_691_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = coreml_update_state_113)[name = string("op_691_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_694_axis_0 = const()[name = string("op_694_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_1 = split(axis = var_694_axis_0, split_sizes = tile_1, x = var_691_cast_fp16)[name = string("op_694_cast_fp16")];
            tensor<int32, [2]> var_697_split_sizes_0 = const()[name = string("op_697_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_697_axis_0 = const()[name = string("op_697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_697_0, tensor<fp16, [1, 8, 128, 32]> var_697_1 = split(axis = var_697_axis_0, split_sizes = var_697_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_697")];
            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_684_cast_fp16_0, y = var_697_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_700_to_fp16 = const()[name = string("op_700_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_700_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_704 = const()[name = string("op_704"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_7_cast_fp16 = softmax(axis = var_704, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_710_transpose_x_1 = const()[name = string("op_710_transpose_x_1"), val = bool(true)];
            bool var_710_transpose_y_1 = const()[name = string("op_710_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_710_cast_fp16 = matmul(transpose_x = var_710_transpose_x_1, transpose_y = var_710_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_694_cast_fp16_0)[name = string("op_710_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_684_cast_fp16_1, y = var_697_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_712_to_fp16 = const()[name = string("op_712_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_712_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_716 = const()[name = string("op_716"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_15_cast_fp16 = softmax(axis = var_716, 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_694_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_724 = const()[name = string("op_724"), 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_724, interleave = attn_output_3_interleave_0, values = (var_710_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_728_perm_0 = const()[name = string("op_728_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_728_cast_fp16 = transpose(perm = var_728_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_219")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_728_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(459814464)))];
            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_761_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_761_cast_fp16")];
            int32 var_759 = const()[name = string("op_759"), 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_759, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_761_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(468203136)))];
            fp16 var_771_to_fp16 = const()[name = string("op_771_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_771_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_782_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = out_3_cast_fp16)[name = string("op_782_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468211392)))];
            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_to_fp16, x = var_782_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_799_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_799_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493377280)))];
            tensor<int32, [2]> var_805_strides_0 = const()[name = string("op_805_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_805_pad_type_0 = const()[name = string("op_805_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_805_pad_0 = const()[name = string("op_805_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_805_dilations_0 = const()[name = string("op_805_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_805_groups_0 = const()[name = string("op_805_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_805_cast_fp16 = conv(dilations = var_805_dilations_0, groups = var_805_groups_0, pad = var_805_pad_0, pad_type = var_805_pad_type_0, strides = var_805_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_782_cast_fp16_0)[name = string("op_805_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_9_cast_fp16 = mul(x = var_799_cast_fp16, y = var_805_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518543168)))];
            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_to_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_823_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_823_cast_fp16")];
            int32 var_821 = const()[name = string("op_821"), 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_821, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_823_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(543709056)))];
            fp16 var_833_to_fp16 = const()[name = string("op_833_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_833_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_844_split_sizes_0 = const()[name = string("op_844_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_844_axis_0 = const()[name = string("op_844_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_844_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_844_cast_fp16_1 = split(axis = var_844_axis_0, split_sizes = var_844_split_sizes_0, x = out_5_cast_fp16)[name = string("op_844_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543717312)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552105984)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            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_cast_fp16, x = var_844_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_901_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_901_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_908_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_908_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_912_cast_fp16 = mul(x = x_11_cast_fp16, y = var_452_cast_fp16)[name = string("op_912_cast_fp16")];
            tensor<int32, [2]> var_913_split_sizes_0 = const()[name = string("op_913_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_913_axis_0 = const()[name = string("op_913_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_913_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_913_cast_fp16_1 = split(axis = var_913_axis_0, split_sizes = var_913_split_sizes_0, x = x_11_cast_fp16)[name = string("op_913_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_915_cast_fp16 = mul(x = var_913_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_915_cast_fp16")];
            int32 var_917 = const()[name = string("op_917"), val = int32(-2)];
            bool var_918_interleave_0 = const()[name = string("op_918_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_918_cast_fp16 = concat(axis = var_917, interleave = var_918_interleave_0, values = (var_915_cast_fp16, var_913_cast_fp16_0))[name = string("op_918_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_919_cast_fp16 = mul(x = var_918_cast_fp16, y = var_459_cast_fp16)[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_9_cast_fp16 = add(x = var_912_cast_fp16, y = var_919_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_925_cast_fp16 = mul(x = var_901_cast_fp16, y = var_452_cast_fp16)[name = string("op_925_cast_fp16")];
            tensor<int32, [2]> var_926_split_sizes_0 = const()[name = string("op_926_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_926_axis_0 = const()[name = string("op_926_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_926_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_926_cast_fp16_1 = split(axis = var_926_axis_0, split_sizes = var_926_split_sizes_0, x = var_901_cast_fp16)[name = string("op_926_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_928_cast_fp16 = mul(x = var_926_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_928_cast_fp16")];
            int32 var_930 = const()[name = string("op_930"), val = int32(-2)];
            bool var_931_interleave_0 = const()[name = string("op_931_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_931_cast_fp16 = concat(axis = var_930, interleave = var_931_interleave_0, values = (var_928_cast_fp16, var_926_cast_fp16_0))[name = string("op_931_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_932_cast_fp16 = mul(x = var_931_cast_fp16, y = var_459_cast_fp16)[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_15_cast_fp16 = add(x = var_925_cast_fp16, y = var_932_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_218")];
            tensor<fp16, [14, 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_112)[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_114_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_114 = read_state(input = key_cache)[name = string("coreml_update_state_114")];
            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_908_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [14, 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_113)[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_115_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_115 = read_state(input = value_cache)[name = string("coreml_update_state_115")];
            tensor<int32, [4]> var_1002_begin_0 = const()[name = string("op_1002_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1002_end_0 = const()[name = string("op_1002_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1002_end_mask_0 = const()[name = string("op_1002_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1002_cast_fp16 = slice_by_index(begin = var_1002_begin_0, end = var_1002_end_0, end_mask = var_1002_end_mask_0, x = coreml_update_state_114)[name = string("op_1002_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1005_axis_0 = const()[name = string("op_1005_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_1 = split(axis = var_1005_axis_0, split_sizes = tile_2, x = var_1002_cast_fp16)[name = string("op_1005_cast_fp16")];
            tensor<int32, [4]> var_1012_begin_0 = const()[name = string("op_1012_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1012_end_0 = const()[name = string("op_1012_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1012_end_mask_0 = const()[name = string("op_1012_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1012_cast_fp16 = slice_by_index(begin = var_1012_begin_0, end = var_1012_end_0, end_mask = var_1012_end_mask_0, x = coreml_update_state_115)[name = string("op_1012_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1015_axis_0 = const()[name = string("op_1015_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_1 = split(axis = var_1015_axis_0, split_sizes = tile_3, x = var_1012_cast_fp16)[name = string("op_1015_cast_fp16")];
            tensor<int32, [2]> var_1018_split_sizes_0 = const()[name = string("op_1018_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1018_axis_0 = const()[name = string("op_1018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1018_0, tensor<fp16, [1, 8, 128, 32]> var_1018_1 = split(axis = var_1018_axis_0, split_sizes = var_1018_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1018")];
            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_1005_cast_fp16_0, y = var_1018_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1021_to_fp16 = const()[name = string("op_1021_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_1021_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_1025 = const()[name = string("op_1025"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_23_cast_fp16 = softmax(axis = var_1025, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1031_transpose_x_1 = const()[name = string("op_1031_transpose_x_1"), val = bool(true)];
            bool var_1031_transpose_y_1 = const()[name = string("op_1031_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1031_cast_fp16 = matmul(transpose_x = var_1031_transpose_x_1, transpose_y = var_1031_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1015_cast_fp16_0)[name = string("op_1031_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_1005_cast_fp16_1, y = var_1018_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1033_to_fp16 = const()[name = string("op_1033_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_1033_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_1037 = const()[name = string("op_1037"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_31_cast_fp16 = softmax(axis = var_1037, 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_1015_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1045 = const()[name = string("op_1045"), 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_1045, interleave = attn_output_11_interleave_0, values = (var_1031_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1049_perm_0 = const()[name = string("op_1049_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_1049_cast_fp16 = transpose(perm = var_1049_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_216")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1049_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553154624)))];
            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_to_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_1082_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1082_cast_fp16")];
            int32 var_1080 = const()[name = string("op_1080"), 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_1080, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1082_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(561543296)))];
            fp16 var_1092_to_fp16 = const()[name = string("op_1092_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_1092_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1103_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1103_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561551552)))];
            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_to_fp16, x = var_1103_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_1120_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [2]> var_1126_strides_0 = const()[name = string("op_1126_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1126_pad_type_0 = const()[name = string("op_1126_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1126_pad_0 = const()[name = string("op_1126_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1126_dilations_0 = const()[name = string("op_1126_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1126_groups_0 = const()[name = string("op_1126_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_1126_cast_fp16 = conv(dilations = var_1126_dilations_0, groups = var_1126_groups_0, pad = var_1126_pad_0, pad_type = var_1126_pad_type_0, strides = var_1126_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1103_cast_fp16_0)[name = string("op_1126_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_19_cast_fp16 = mul(x = var_1120_cast_fp16, y = var_1126_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586717440)))];
            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_to_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_1144_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1144_cast_fp16")];
            int32 var_1142 = const()[name = string("op_1142"), 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_1142, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1144_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(611883328)))];
            fp16 var_1154_to_fp16 = const()[name = string("op_1154_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_1154_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1165_split_sizes_0 = const()[name = string("op_1165_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1165_axis_0 = const()[name = string("op_1165_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1165_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1165_cast_fp16_1 = split(axis = var_1165_axis_0, split_sizes = var_1165_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1165_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611891584)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620280256)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            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_cast_fp16, x = var_1165_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_1222_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1222_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_1229_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1229_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1233_cast_fp16 = mul(x = x_21_cast_fp16, y = var_452_cast_fp16)[name = string("op_1233_cast_fp16")];
            tensor<int32, [2]> var_1234_split_sizes_0 = const()[name = string("op_1234_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1234_axis_0 = const()[name = string("op_1234_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_1234_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_1234_cast_fp16_1 = split(axis = var_1234_axis_0, split_sizes = var_1234_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1234_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_1236_cast_fp16 = mul(x = var_1234_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1236_cast_fp16")];
            int32 var_1238 = const()[name = string("op_1238"), val = int32(-2)];
            bool var_1239_interleave_0 = const()[name = string("op_1239_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_1239_cast_fp16 = concat(axis = var_1238, interleave = var_1239_interleave_0, values = (var_1236_cast_fp16, var_1234_cast_fp16_0))[name = string("op_1239_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1240_cast_fp16 = mul(x = var_1239_cast_fp16, y = var_459_cast_fp16)[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_15_cast_fp16 = add(x = var_1233_cast_fp16, y = var_1240_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1246_cast_fp16 = mul(x = var_1222_cast_fp16, y = var_452_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<int32, [2]> var_1247_split_sizes_0 = const()[name = string("op_1247_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1247_axis_0 = const()[name = string("op_1247_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_1247_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_1247_cast_fp16_1 = split(axis = var_1247_axis_0, split_sizes = var_1247_split_sizes_0, x = var_1222_cast_fp16)[name = string("op_1247_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_1249_cast_fp16 = mul(x = var_1247_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1249_cast_fp16")];
            int32 var_1251 = const()[name = string("op_1251"), val = int32(-2)];
            bool var_1252_interleave_0 = const()[name = string("op_1252_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_1252_cast_fp16 = concat(axis = var_1251, interleave = var_1252_interleave_0, values = (var_1249_cast_fp16, var_1247_cast_fp16_0))[name = string("op_1252_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1253_cast_fp16 = mul(x = var_1252_cast_fp16, y = var_459_cast_fp16)[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_25_cast_fp16 = add(x = var_1246_cast_fp16, y = var_1253_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_215")];
            tensor<fp16, [14, 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_114)[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_116_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_116 = read_state(input = key_cache)[name = string("coreml_update_state_116")];
            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_1229_cast_fp16)[name = string("transpose_214")];
            tensor<fp16, [14, 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_115)[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_117_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_117 = read_state(input = value_cache)[name = string("coreml_update_state_117")];
            tensor<int32, [4]> var_1323_begin_0 = const()[name = string("op_1323_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1323_end_0 = const()[name = string("op_1323_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1323_end_mask_0 = const()[name = string("op_1323_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1323_cast_fp16 = slice_by_index(begin = var_1323_begin_0, end = var_1323_end_0, end_mask = var_1323_end_mask_0, x = coreml_update_state_116)[name = string("op_1323_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1326_axis_0 = const()[name = string("op_1326_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_1 = split(axis = var_1326_axis_0, split_sizes = tile_4, x = var_1323_cast_fp16)[name = string("op_1326_cast_fp16")];
            tensor<int32, [4]> var_1333_begin_0 = const()[name = string("op_1333_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1333_end_0 = const()[name = string("op_1333_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1333_end_mask_0 = const()[name = string("op_1333_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1333_cast_fp16 = slice_by_index(begin = var_1333_begin_0, end = var_1333_end_0, end_mask = var_1333_end_mask_0, x = coreml_update_state_117)[name = string("op_1333_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1336_axis_0 = const()[name = string("op_1336_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_1 = split(axis = var_1336_axis_0, split_sizes = tile_5, x = var_1333_cast_fp16)[name = string("op_1336_cast_fp16")];
            tensor<int32, [2]> var_1339_split_sizes_0 = const()[name = string("op_1339_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1339_axis_0 = const()[name = string("op_1339_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1339_0, tensor<fp16, [1, 8, 128, 32]> var_1339_1 = split(axis = var_1339_axis_0, split_sizes = var_1339_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1339")];
            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_1326_cast_fp16_0, y = var_1339_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1342_to_fp16 = const()[name = string("op_1342_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_1342_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_1346 = const()[name = string("op_1346"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_39_cast_fp16 = softmax(axis = var_1346, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1352_transpose_x_1 = const()[name = string("op_1352_transpose_x_1"), val = bool(true)];
            bool var_1352_transpose_y_1 = const()[name = string("op_1352_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1352_cast_fp16 = matmul(transpose_x = var_1352_transpose_x_1, transpose_y = var_1352_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1336_cast_fp16_0)[name = string("op_1352_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_1326_cast_fp16_1, y = var_1339_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1354_to_fp16 = const()[name = string("op_1354_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_1354_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_1358 = const()[name = string("op_1358"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_47_cast_fp16 = softmax(axis = var_1358, 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_1336_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1366 = const()[name = string("op_1366"), 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_1366, interleave = attn_output_19_interleave_0, values = (var_1352_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1370_perm_0 = const()[name = string("op_1370_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_1370_cast_fp16 = transpose(perm = var_1370_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_213")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1370_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621328896)))];
            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_to_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_1403_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1403_cast_fp16")];
            int32 var_1401 = const()[name = string("op_1401"), 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_1401, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1403_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(629717568)))];
            fp16 var_1413_to_fp16 = const()[name = string("op_1413_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_1413_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1424_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1424_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629725824)))];
            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_to_fp16, x = var_1424_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_1441_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1441_cast_fp16")];
            tensor<int32, [2]> var_1447_strides_0 = const()[name = string("op_1447_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1447_pad_type_0 = const()[name = string("op_1447_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1447_pad_0 = const()[name = string("op_1447_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1447_dilations_0 = const()[name = string("op_1447_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1447_groups_0 = const()[name = string("op_1447_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_1447_cast_fp16 = conv(dilations = var_1447_dilations_0, groups = var_1447_groups_0, pad = var_1447_pad_0, pad_type = var_1447_pad_type_0, strides = var_1447_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1424_cast_fp16_0)[name = string("op_1447_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_29_cast_fp16 = mul(x = var_1441_cast_fp16, y = var_1447_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654891712)))];
            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_to_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_1465_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1465_cast_fp16")];
            int32 var_1463 = const()[name = string("op_1463"), 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_1463, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1465_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(680057600)))];
            fp16 var_1475_to_fp16 = const()[name = string("op_1475_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_1475_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1486_split_sizes_0 = const()[name = string("op_1486_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1486_axis_0 = const()[name = string("op_1486_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1486_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1486_cast_fp16_1 = split(axis = var_1486_axis_0, split_sizes = var_1486_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1486_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680065856)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688454528)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            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_cast_fp16, x = var_1486_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_1543_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1543_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_1550_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1554_cast_fp16 = mul(x = x_31_cast_fp16, y = var_452_cast_fp16)[name = string("op_1554_cast_fp16")];
            tensor<int32, [2]> var_1555_split_sizes_0 = const()[name = string("op_1555_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1555_axis_0 = const()[name = string("op_1555_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_1555_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_1555_cast_fp16_1 = split(axis = var_1555_axis_0, split_sizes = var_1555_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1555_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_1557_cast_fp16 = mul(x = var_1555_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1557_cast_fp16")];
            int32 var_1559 = const()[name = string("op_1559"), val = int32(-2)];
            bool var_1560_interleave_0 = const()[name = string("op_1560_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_1560_cast_fp16 = concat(axis = var_1559, interleave = var_1560_interleave_0, values = (var_1557_cast_fp16, var_1555_cast_fp16_0))[name = string("op_1560_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1561_cast_fp16 = mul(x = var_1560_cast_fp16, y = var_459_cast_fp16)[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_21_cast_fp16 = add(x = var_1554_cast_fp16, y = var_1561_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1567_cast_fp16 = mul(x = var_1543_cast_fp16, y = var_452_cast_fp16)[name = string("op_1567_cast_fp16")];
            tensor<int32, [2]> var_1568_split_sizes_0 = const()[name = string("op_1568_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1568_axis_0 = const()[name = string("op_1568_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_1568_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_1568_cast_fp16_1 = split(axis = var_1568_axis_0, split_sizes = var_1568_split_sizes_0, x = var_1543_cast_fp16)[name = string("op_1568_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_1570_cast_fp16 = mul(x = var_1568_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1570_cast_fp16")];
            int32 var_1572 = const()[name = string("op_1572"), val = int32(-2)];
            bool var_1573_interleave_0 = const()[name = string("op_1573_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_1573_cast_fp16 = concat(axis = var_1572, interleave = var_1573_interleave_0, values = (var_1570_cast_fp16, var_1568_cast_fp16_0))[name = string("op_1573_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1574_cast_fp16 = mul(x = var_1573_cast_fp16, y = var_459_cast_fp16)[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_35_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1574_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_212")];
            tensor<fp16, [14, 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_116)[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_118_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_118 = read_state(input = key_cache)[name = string("coreml_update_state_118")];
            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_1550_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [14, 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_117)[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_119_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_119 = read_state(input = value_cache)[name = string("coreml_update_state_119")];
            tensor<int32, [4]> var_1644_begin_0 = const()[name = string("op_1644_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1644_end_0 = const()[name = string("op_1644_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1644_end_mask_0 = const()[name = string("op_1644_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1644_cast_fp16 = slice_by_index(begin = var_1644_begin_0, end = var_1644_end_0, end_mask = var_1644_end_mask_0, x = coreml_update_state_118)[name = string("op_1644_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1647_axis_0 = const()[name = string("op_1647_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_1 = split(axis = var_1647_axis_0, split_sizes = tile_6, x = var_1644_cast_fp16)[name = string("op_1647_cast_fp16")];
            tensor<int32, [4]> var_1654_begin_0 = const()[name = string("op_1654_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1654_end_0 = const()[name = string("op_1654_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1654_end_mask_0 = const()[name = string("op_1654_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1654_cast_fp16 = slice_by_index(begin = var_1654_begin_0, end = var_1654_end_0, end_mask = var_1654_end_mask_0, x = coreml_update_state_119)[name = string("op_1654_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1657_axis_0 = const()[name = string("op_1657_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_1 = split(axis = var_1657_axis_0, split_sizes = tile_7, x = var_1654_cast_fp16)[name = string("op_1657_cast_fp16")];
            tensor<int32, [2]> var_1660_split_sizes_0 = const()[name = string("op_1660_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1660_axis_0 = const()[name = string("op_1660_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1660_0, tensor<fp16, [1, 8, 128, 32]> var_1660_1 = split(axis = var_1660_axis_0, split_sizes = var_1660_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1660")];
            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_1647_cast_fp16_0, y = var_1660_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1663_to_fp16 = const()[name = string("op_1663_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_1663_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_1667 = const()[name = string("op_1667"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_55_cast_fp16 = softmax(axis = var_1667, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1673_transpose_x_1 = const()[name = string("op_1673_transpose_x_1"), val = bool(true)];
            bool var_1673_transpose_y_1 = const()[name = string("op_1673_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1673_cast_fp16 = matmul(transpose_x = var_1673_transpose_x_1, transpose_y = var_1673_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1657_cast_fp16_0)[name = string("op_1673_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_1647_cast_fp16_1, y = var_1660_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1675_to_fp16 = const()[name = string("op_1675_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_1675_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_1679 = const()[name = string("op_1679"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_63_cast_fp16 = softmax(axis = var_1679, 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_1657_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1687 = const()[name = string("op_1687"), 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_1687, interleave = attn_output_27_interleave_0, values = (var_1673_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1691_perm_0 = const()[name = string("op_1691_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_1691_cast_fp16 = transpose(perm = var_1691_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_210")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1691_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            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_cast_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_1724_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1724_cast_fp16")];
            int32 var_1722 = const()[name = string("op_1722"), 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_1722, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1724_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(689503168)))];
            fp16 var_1734_to_fp16 = const()[name = string("op_1734_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_1734_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1745_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1745_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689511424)))];
            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_to_fp16, x = var_1745_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_1762_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1762_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(714677312)))];
            tensor<int32, [2]> var_1768_strides_0 = const()[name = string("op_1768_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1768_pad_type_0 = const()[name = string("op_1768_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1768_pad_0 = const()[name = string("op_1768_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1768_dilations_0 = const()[name = string("op_1768_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1768_groups_0 = const()[name = string("op_1768_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_1768_cast_fp16 = conv(dilations = var_1768_dilations_0, groups = var_1768_groups_0, pad = var_1768_pad_0, pad_type = var_1768_pad_type_0, strides = var_1768_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1745_cast_fp16_0)[name = string("op_1768_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_39_cast_fp16 = mul(x = var_1762_cast_fp16, y = var_1768_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_1786_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1786_cast_fp16")];
            int32 var_1784 = const()[name = string("op_1784"), 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_1784, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1786_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(739843200)))];
            fp16 var_1796_to_fp16 = const()[name = string("op_1796_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_1796_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1807_split_sizes_0 = const()[name = string("op_1807_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1807_axis_0 = const()[name = string("op_1807_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_1807_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_1807_cast_fp16_1 = split(axis = var_1807_axis_0, split_sizes = var_1807_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739851456)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(748240128)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            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_cast_fp16, x = var_1807_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_1864_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1864_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_1871_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1871_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1875_cast_fp16 = mul(x = x_41_cast_fp16, y = var_452_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [2]> var_1876_split_sizes_0 = const()[name = string("op_1876_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1876_axis_0 = const()[name = string("op_1876_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_1876_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_1876_cast_fp16_1 = split(axis = var_1876_axis_0, split_sizes = var_1876_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1876_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_1878_cast_fp16 = mul(x = var_1876_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1878_cast_fp16")];
            int32 var_1880 = const()[name = string("op_1880"), val = int32(-2)];
            bool var_1881_interleave_0 = const()[name = string("op_1881_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_1881_cast_fp16 = concat(axis = var_1880, interleave = var_1881_interleave_0, values = (var_1878_cast_fp16, var_1876_cast_fp16_0))[name = string("op_1881_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_1882_cast_fp16 = mul(x = var_1881_cast_fp16, y = var_459_cast_fp16)[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_27_cast_fp16 = add(x = var_1875_cast_fp16, y = var_1882_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1888_cast_fp16 = mul(x = var_1864_cast_fp16, y = var_452_cast_fp16)[name = string("op_1888_cast_fp16")];
            tensor<int32, [2]> var_1889_split_sizes_0 = const()[name = string("op_1889_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1889_axis_0 = const()[name = string("op_1889_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_1889_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_1889_cast_fp16_1 = split(axis = var_1889_axis_0, split_sizes = var_1889_split_sizes_0, x = var_1864_cast_fp16)[name = string("op_1889_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_1891_cast_fp16 = mul(x = var_1889_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1891_cast_fp16")];
            int32 var_1893 = const()[name = string("op_1893"), val = int32(-2)];
            bool var_1894_interleave_0 = const()[name = string("op_1894_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_1894_cast_fp16 = concat(axis = var_1893, interleave = var_1894_interleave_0, values = (var_1891_cast_fp16, var_1889_cast_fp16_0))[name = string("op_1894_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_1895_cast_fp16 = mul(x = var_1894_cast_fp16, y = var_459_cast_fp16)[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_45_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1895_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_209")];
            tensor<fp16, [14, 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_118)[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_120_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_120 = read_state(input = key_cache)[name = string("coreml_update_state_120")];
            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_1871_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [14, 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_119)[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_121_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_121 = read_state(input = value_cache)[name = string("coreml_update_state_121")];
            tensor<int32, [4]> var_1965_begin_0 = const()[name = string("op_1965_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1965_end_0 = const()[name = string("op_1965_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1965_end_mask_0 = const()[name = string("op_1965_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1965_cast_fp16 = slice_by_index(begin = var_1965_begin_0, end = var_1965_end_0, end_mask = var_1965_end_mask_0, x = coreml_update_state_120)[name = string("op_1965_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1968_axis_0 = const()[name = string("op_1968_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_1 = split(axis = var_1968_axis_0, split_sizes = tile_8, x = var_1965_cast_fp16)[name = string("op_1968_cast_fp16")];
            tensor<int32, [4]> var_1975_begin_0 = const()[name = string("op_1975_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1975_end_0 = const()[name = string("op_1975_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1975_end_mask_0 = const()[name = string("op_1975_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1975_cast_fp16 = slice_by_index(begin = var_1975_begin_0, end = var_1975_end_0, end_mask = var_1975_end_mask_0, x = coreml_update_state_121)[name = string("op_1975_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1978_axis_0 = const()[name = string("op_1978_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_1 = split(axis = var_1978_axis_0, split_sizes = tile_9, x = var_1975_cast_fp16)[name = string("op_1978_cast_fp16")];
            tensor<int32, [2]> var_1981_split_sizes_0 = const()[name = string("op_1981_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1981_axis_0 = const()[name = string("op_1981_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_1981_0, tensor<fp16, [1, 8, 128, 32]> var_1981_1 = split(axis = var_1981_axis_0, split_sizes = var_1981_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1981")];
            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_1968_cast_fp16_0, y = var_1981_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1984_to_fp16 = const()[name = string("op_1984_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_1984_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_1988 = const()[name = string("op_1988"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_71_cast_fp16 = softmax(axis = var_1988, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1994_transpose_x_1 = const()[name = string("op_1994_transpose_x_1"), val = bool(true)];
            bool var_1994_transpose_y_1 = const()[name = string("op_1994_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1994_cast_fp16 = matmul(transpose_x = var_1994_transpose_x_1, transpose_y = var_1994_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1978_cast_fp16_0)[name = string("op_1994_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_1968_cast_fp16_1, y = var_1981_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1996_to_fp16 = const()[name = string("op_1996_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_1996_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_2000 = const()[name = string("op_2000"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_79_cast_fp16 = softmax(axis = var_2000, 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_1978_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2008 = const()[name = string("op_2008"), 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_2008, interleave = attn_output_35_interleave_0, values = (var_1994_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2012_perm_0 = const()[name = string("op_2012_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_2012_cast_fp16 = transpose(perm = var_2012_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_207")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2012_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_2045_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2045_cast_fp16")];
            int32 var_2043 = const()[name = string("op_2043"), 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_2043, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2045_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(749288768)))];
            fp16 var_2055_to_fp16 = const()[name = string("op_2055_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_2055_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2066_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2066_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_2066_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_2083_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2083_cast_fp16")];
            tensor<int32, [2]> var_2089_strides_0 = const()[name = string("op_2089_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2089_pad_type_0 = const()[name = string("op_2089_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2089_pad_0 = const()[name = string("op_2089_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2089_dilations_0 = const()[name = string("op_2089_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2089_groups_0 = const()[name = string("op_2089_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_2089_cast_fp16 = conv(dilations = var_2089_dilations_0, groups = var_2089_groups_0, pad = var_2089_pad_0, pad_type = var_2089_pad_type_0, strides = var_2089_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2066_cast_fp16_0)[name = string("op_2089_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_49_cast_fp16 = mul(x = var_2083_cast_fp16, y = var_2089_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_2107_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2107_cast_fp16")];
            int32 var_2105 = const()[name = string("op_2105"), 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_2105, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2107_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(749297024)))];
            fp16 var_2117_to_fp16 = const()[name = string("op_2117_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_2117_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2128_split_sizes_0 = const()[name = string("op_2128_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2128_axis_0 = const()[name = string("op_2128_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2128_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2128_cast_fp16_1 = split(axis = var_2128_axis_0, split_sizes = var_2128_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749305280)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(757693952)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            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_cast_fp16, x = var_2128_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_2185_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2185_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_2192_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2196_cast_fp16 = mul(x = x_51_cast_fp16, y = var_452_cast_fp16)[name = string("op_2196_cast_fp16")];
            tensor<int32, [2]> var_2197_split_sizes_0 = const()[name = string("op_2197_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2197_axis_0 = const()[name = string("op_2197_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_2197_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_2197_cast_fp16_1 = split(axis = var_2197_axis_0, split_sizes = var_2197_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2197_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_2199_cast_fp16 = mul(x = var_2197_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2199_cast_fp16")];
            int32 var_2201 = const()[name = string("op_2201"), val = int32(-2)];
            bool var_2202_interleave_0 = const()[name = string("op_2202_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_2202_cast_fp16 = concat(axis = var_2201, interleave = var_2202_interleave_0, values = (var_2199_cast_fp16, var_2197_cast_fp16_0))[name = string("op_2202_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2203_cast_fp16 = mul(x = var_2202_cast_fp16, y = var_459_cast_fp16)[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_33_cast_fp16 = add(x = var_2196_cast_fp16, y = var_2203_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2209_cast_fp16 = mul(x = var_2185_cast_fp16, y = var_452_cast_fp16)[name = string("op_2209_cast_fp16")];
            tensor<int32, [2]> var_2210_split_sizes_0 = const()[name = string("op_2210_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2210_axis_0 = const()[name = string("op_2210_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_2210_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_2210_cast_fp16_1 = split(axis = var_2210_axis_0, split_sizes = var_2210_split_sizes_0, x = var_2185_cast_fp16)[name = string("op_2210_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_2212_cast_fp16 = mul(x = var_2210_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2212_cast_fp16")];
            int32 var_2214 = const()[name = string("op_2214"), val = int32(-2)];
            bool var_2215_interleave_0 = const()[name = string("op_2215_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_2215_cast_fp16 = concat(axis = var_2214, interleave = var_2215_interleave_0, values = (var_2212_cast_fp16, var_2210_cast_fp16_0))[name = string("op_2215_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2216_cast_fp16 = mul(x = var_2215_cast_fp16, y = var_459_cast_fp16)[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_55_cast_fp16 = add(x = var_2209_cast_fp16, y = var_2216_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_206")];
            tensor<fp16, [14, 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_120)[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_122_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_122 = read_state(input = key_cache)[name = string("coreml_update_state_122")];
            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_2192_cast_fp16)[name = string("transpose_205")];
            tensor<fp16, [14, 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_121)[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_123_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_123 = read_state(input = value_cache)[name = string("coreml_update_state_123")];
            tensor<int32, [4]> var_2286_begin_0 = const()[name = string("op_2286_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2286_end_0 = const()[name = string("op_2286_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2286_end_mask_0 = const()[name = string("op_2286_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2286_cast_fp16 = slice_by_index(begin = var_2286_begin_0, end = var_2286_end_0, end_mask = var_2286_end_mask_0, x = coreml_update_state_122)[name = string("op_2286_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2289_axis_0 = const()[name = string("op_2289_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_1 = split(axis = var_2289_axis_0, split_sizes = tile_10, x = var_2286_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> var_2296_begin_0 = const()[name = string("op_2296_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2296_end_0 = const()[name = string("op_2296_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2296_end_mask_0 = const()[name = string("op_2296_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2296_cast_fp16 = slice_by_index(begin = var_2296_begin_0, end = var_2296_end_0, end_mask = var_2296_end_mask_0, x = coreml_update_state_123)[name = string("op_2296_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2299_axis_0 = const()[name = string("op_2299_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_1 = split(axis = var_2299_axis_0, split_sizes = tile_11, x = var_2296_cast_fp16)[name = string("op_2299_cast_fp16")];
            tensor<int32, [2]> var_2302_split_sizes_0 = const()[name = string("op_2302_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2302_axis_0 = const()[name = string("op_2302_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_2302_0, tensor<fp16, [1, 8, 128, 32]> var_2302_1 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2302")];
            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_2289_cast_fp16_0, y = var_2302_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2305_to_fp16 = const()[name = string("op_2305_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_2305_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_2309 = const()[name = string("op_2309"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_87_cast_fp16 = softmax(axis = var_2309, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2315_transpose_x_1 = const()[name = string("op_2315_transpose_x_1"), val = bool(true)];
            bool var_2315_transpose_y_1 = const()[name = string("op_2315_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2315_cast_fp16 = matmul(transpose_x = var_2315_transpose_x_1, transpose_y = var_2315_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2299_cast_fp16_0)[name = string("op_2315_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_2289_cast_fp16_1, y = var_2302_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2317_to_fp16 = const()[name = string("op_2317_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_2317_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_2321 = const()[name = string("op_2321"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_95_cast_fp16 = softmax(axis = var_2321, 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_2299_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2329 = const()[name = string("op_2329"), 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_2329, interleave = attn_output_43_interleave_0, values = (var_2315_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2333_perm_0 = const()[name = string("op_2333_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_2333_cast_fp16 = transpose(perm = var_2333_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_204")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2333_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_2366_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2366_cast_fp16")];
            int32 var_2364 = const()[name = string("op_2364"), 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_2364, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2366_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(758742592)))];
            fp16 var_2376_to_fp16 = const()[name = string("op_2376_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_2376_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2387_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2387_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(758750848)))];
            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_to_fp16, x = var_2387_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_2404_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [2]> var_2410_strides_0 = const()[name = string("op_2410_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2410_pad_type_0 = const()[name = string("op_2410_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2410_pad_0 = const()[name = string("op_2410_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2410_dilations_0 = const()[name = string("op_2410_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2410_groups_0 = const()[name = string("op_2410_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_2410_cast_fp16 = conv(dilations = var_2410_dilations_0, groups = var_2410_groups_0, pad = var_2410_pad_0, pad_type = var_2410_pad_type_0, strides = var_2410_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2387_cast_fp16_0)[name = string("op_2410_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_59_cast_fp16 = mul(x = var_2404_cast_fp16, y = var_2410_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_2428_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2428_cast_fp16")];
            int32 var_2426 = const()[name = string("op_2426"), 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_2426, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2428_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(783916736)))];
            fp16 var_2438_to_fp16 = const()[name = string("op_2438_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_2438_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2449_split_sizes_0 = const()[name = string("op_2449_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2449_axis_0 = const()[name = string("op_2449_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2449_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2449_cast_fp16_1 = split(axis = var_2449_axis_0, split_sizes = var_2449_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2449_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(783924992)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792313664)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            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_cast_fp16, x = var_2449_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_2506_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2506_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_2513_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2517_cast_fp16 = mul(x = x_61_cast_fp16, y = var_452_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<int32, [2]> var_2518_split_sizes_0 = const()[name = string("op_2518_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2518_axis_0 = const()[name = string("op_2518_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_2518_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_2518_cast_fp16_1 = split(axis = var_2518_axis_0, split_sizes = var_2518_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2518_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_2520_cast_fp16 = mul(x = var_2518_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2520_cast_fp16")];
            int32 var_2522 = const()[name = string("op_2522"), val = int32(-2)];
            bool var_2523_interleave_0 = const()[name = string("op_2523_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_2523_cast_fp16 = concat(axis = var_2522, interleave = var_2523_interleave_0, values = (var_2520_cast_fp16, var_2518_cast_fp16_0))[name = string("op_2523_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2524_cast_fp16 = mul(x = var_2523_cast_fp16, y = var_459_cast_fp16)[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_39_cast_fp16 = add(x = var_2517_cast_fp16, y = var_2524_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2530_cast_fp16 = mul(x = var_2506_cast_fp16, y = var_452_cast_fp16)[name = string("op_2530_cast_fp16")];
            tensor<int32, [2]> var_2531_split_sizes_0 = const()[name = string("op_2531_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2531_axis_0 = const()[name = string("op_2531_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_2531_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_2531_cast_fp16_1 = split(axis = var_2531_axis_0, split_sizes = var_2531_split_sizes_0, x = var_2506_cast_fp16)[name = string("op_2531_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_2533_cast_fp16 = mul(x = var_2531_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2533_cast_fp16")];
            int32 var_2535 = const()[name = string("op_2535"), val = int32(-2)];
            bool var_2536_interleave_0 = const()[name = string("op_2536_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_2536_cast_fp16 = concat(axis = var_2535, interleave = var_2536_interleave_0, values = (var_2533_cast_fp16, var_2531_cast_fp16_0))[name = string("op_2536_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2537_cast_fp16 = mul(x = var_2536_cast_fp16, y = var_459_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_65_cast_fp16 = add(x = var_2530_cast_fp16, y = var_2537_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_203")];
            tensor<fp16, [14, 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_122)[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_124_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_124 = read_state(input = key_cache)[name = string("coreml_update_state_124")];
            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_2513_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [14, 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_123)[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_125_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_125 = read_state(input = value_cache)[name = string("coreml_update_state_125")];
            tensor<int32, [4]> var_2607_begin_0 = const()[name = string("op_2607_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2607_end_0 = const()[name = string("op_2607_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2607_end_mask_0 = const()[name = string("op_2607_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2607_cast_fp16 = slice_by_index(begin = var_2607_begin_0, end = var_2607_end_0, end_mask = var_2607_end_mask_0, x = coreml_update_state_124)[name = string("op_2607_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2610_axis_0 = const()[name = string("op_2610_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_1 = split(axis = var_2610_axis_0, split_sizes = tile_12, x = var_2607_cast_fp16)[name = string("op_2610_cast_fp16")];
            tensor<int32, [4]> var_2617_begin_0 = const()[name = string("op_2617_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2617_end_0 = const()[name = string("op_2617_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2617_end_mask_0 = const()[name = string("op_2617_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2617_cast_fp16 = slice_by_index(begin = var_2617_begin_0, end = var_2617_end_0, end_mask = var_2617_end_mask_0, x = coreml_update_state_125)[name = string("op_2617_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2620_axis_0 = const()[name = string("op_2620_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_1 = split(axis = var_2620_axis_0, split_sizes = tile_13, x = var_2617_cast_fp16)[name = string("op_2620_cast_fp16")];
            tensor<int32, [2]> var_2623_split_sizes_0 = const()[name = string("op_2623_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2623_axis_0 = const()[name = string("op_2623_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_2623_0, tensor<fp16, [1, 8, 128, 32]> var_2623_1 = split(axis = var_2623_axis_0, split_sizes = var_2623_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2623")];
            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_2610_cast_fp16_0, y = var_2623_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2626_to_fp16 = const()[name = string("op_2626_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_2626_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_2630 = const()[name = string("op_2630"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_103_cast_fp16 = softmax(axis = var_2630, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2636_transpose_x_1 = const()[name = string("op_2636_transpose_x_1"), val = bool(true)];
            bool var_2636_transpose_y_1 = const()[name = string("op_2636_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2636_cast_fp16 = matmul(transpose_x = var_2636_transpose_x_1, transpose_y = var_2636_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2620_cast_fp16_0)[name = string("op_2636_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_2610_cast_fp16_1, y = var_2623_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2638_to_fp16 = const()[name = string("op_2638_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_2638_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_2642 = const()[name = string("op_2642"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_111_cast_fp16 = softmax(axis = var_2642, 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_2620_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2650 = const()[name = string("op_2650"), 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_2650, interleave = attn_output_51_interleave_0, values = (var_2636_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2654_perm_0 = const()[name = string("op_2654_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_2654_cast_fp16 = transpose(perm = var_2654_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_201")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2654_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_2687_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2687_cast_fp16")];
            int32 var_2685 = const()[name = string("op_2685"), 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_2685, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2687_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(793362304)))];
            fp16 var_2697_to_fp16 = const()[name = string("op_2697_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_2697_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2708_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2708_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_2708_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_2725_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2725_cast_fp16")];
            tensor<int32, [2]> var_2731_strides_0 = const()[name = string("op_2731_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2731_pad_type_0 = const()[name = string("op_2731_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2731_pad_0 = const()[name = string("op_2731_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2731_dilations_0 = const()[name = string("op_2731_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2731_groups_0 = const()[name = string("op_2731_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_2731_cast_fp16 = conv(dilations = var_2731_dilations_0, groups = var_2731_groups_0, pad = var_2731_pad_0, pad_type = var_2731_pad_type_0, strides = var_2731_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2708_cast_fp16_0)[name = string("op_2731_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_69_cast_fp16 = mul(x = var_2725_cast_fp16, y = var_2731_cast_fp16)[name = string("x_69_cast_fp16")];
            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_cast_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_2749_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2749_cast_fp16")];
            int32 var_2747 = const()[name = string("op_2747"), 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_2747, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2749_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(793370560)))];
            fp16 var_2759_to_fp16 = const()[name = string("op_2759_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_2759_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2770_split_sizes_0 = const()[name = string("op_2770_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2770_axis_0 = const()[name = string("op_2770_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_2770_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_2770_cast_fp16_1 = split(axis = var_2770_axis_0, split_sizes = var_2770_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2770_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(793378816)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(801767488)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            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_cast_fp16, x = var_2770_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_2827_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2827_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_2834_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2834_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2838_cast_fp16 = mul(x = x_71_cast_fp16, y = var_452_cast_fp16)[name = string("op_2838_cast_fp16")];
            tensor<int32, [2]> var_2839_split_sizes_0 = const()[name = string("op_2839_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2839_axis_0 = const()[name = string("op_2839_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_2839_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_2839_cast_fp16_1 = split(axis = var_2839_axis_0, split_sizes = var_2839_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2839_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_2841_cast_fp16 = mul(x = var_2839_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2841_cast_fp16")];
            int32 var_2843 = const()[name = string("op_2843"), val = int32(-2)];
            bool var_2844_interleave_0 = const()[name = string("op_2844_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_2844_cast_fp16 = concat(axis = var_2843, interleave = var_2844_interleave_0, values = (var_2841_cast_fp16, var_2839_cast_fp16_0))[name = string("op_2844_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_2845_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_459_cast_fp16)[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_45_cast_fp16 = add(x = var_2838_cast_fp16, y = var_2845_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2851_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_452_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [2]> var_2852_split_sizes_0 = const()[name = string("op_2852_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2852_axis_0 = const()[name = string("op_2852_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_2852_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_2852_cast_fp16_1 = split(axis = var_2852_axis_0, split_sizes = var_2852_split_sizes_0, x = var_2827_cast_fp16)[name = string("op_2852_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_2854_cast_fp16 = mul(x = var_2852_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2854_cast_fp16")];
            int32 var_2856 = const()[name = string("op_2856"), val = int32(-2)];
            bool var_2857_interleave_0 = const()[name = string("op_2857_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_2857_cast_fp16 = concat(axis = var_2856, interleave = var_2857_interleave_0, values = (var_2854_cast_fp16, var_2852_cast_fp16_0))[name = string("op_2857_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_2858_cast_fp16 = mul(x = var_2857_cast_fp16, y = var_459_cast_fp16)[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_75_cast_fp16 = add(x = var_2851_cast_fp16, y = var_2858_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_200")];
            tensor<fp16, [14, 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_124)[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_126_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_126 = read_state(input = key_cache)[name = string("coreml_update_state_126")];
            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_2834_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [14, 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_125)[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_127_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_127 = read_state(input = value_cache)[name = string("coreml_update_state_127")];
            tensor<int32, [4]> var_2928_begin_0 = const()[name = string("op_2928_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2928_end_0 = const()[name = string("op_2928_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2928_end_mask_0 = const()[name = string("op_2928_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2928_cast_fp16 = slice_by_index(begin = var_2928_begin_0, end = var_2928_end_0, end_mask = var_2928_end_mask_0, x = coreml_update_state_126)[name = string("op_2928_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2931_axis_0 = const()[name = string("op_2931_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_1 = split(axis = var_2931_axis_0, split_sizes = tile_14, x = var_2928_cast_fp16)[name = string("op_2931_cast_fp16")];
            tensor<int32, [4]> var_2938_begin_0 = const()[name = string("op_2938_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2938_end_0 = const()[name = string("op_2938_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2938_end_mask_0 = const()[name = string("op_2938_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2938_cast_fp16 = slice_by_index(begin = var_2938_begin_0, end = var_2938_end_0, end_mask = var_2938_end_mask_0, x = coreml_update_state_127)[name = string("op_2938_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2941_axis_0 = const()[name = string("op_2941_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_1 = split(axis = var_2941_axis_0, split_sizes = tile_15, x = var_2938_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<int32, [2]> var_2944_split_sizes_0 = const()[name = string("op_2944_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2944_axis_0 = const()[name = string("op_2944_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_2944_0, tensor<fp16, [1, 8, 128, 32]> var_2944_1 = split(axis = var_2944_axis_0, split_sizes = var_2944_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2944")];
            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_2931_cast_fp16_0, y = var_2944_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2947_to_fp16 = const()[name = string("op_2947_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_2947_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_2951 = const()[name = string("op_2951"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_119_cast_fp16 = softmax(axis = var_2951, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2957_transpose_x_1 = const()[name = string("op_2957_transpose_x_1"), val = bool(true)];
            bool var_2957_transpose_y_1 = const()[name = string("op_2957_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2957_cast_fp16 = matmul(transpose_x = var_2957_transpose_x_1, transpose_y = var_2957_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2941_cast_fp16_0)[name = string("op_2957_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_2931_cast_fp16_1, y = var_2944_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2959_to_fp16 = const()[name = string("op_2959_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_2959_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_2963 = const()[name = string("op_2963"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_127_cast_fp16 = softmax(axis = var_2963, 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_2941_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2971 = const()[name = string("op_2971"), 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_2971, interleave = attn_output_59_interleave_0, values = (var_2957_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2975_perm_0 = const()[name = string("op_2975_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_2975_cast_fp16 = transpose(perm = var_2975_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_198")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2975_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_3008_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3008_cast_fp16")];
            int32 var_3006 = const()[name = string("op_3006"), 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_3006, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3008_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(802816128)))];
            fp16 var_3018_to_fp16 = const()[name = string("op_3018_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_3018_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3029_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3029_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_3029_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_3046_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3046_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(802824384)))];
            tensor<int32, [2]> var_3052_strides_0 = const()[name = string("op_3052_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3052_pad_type_0 = const()[name = string("op_3052_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3052_pad_0 = const()[name = string("op_3052_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3052_dilations_0 = const()[name = string("op_3052_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3052_groups_0 = const()[name = string("op_3052_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_3052_cast_fp16 = conv(dilations = var_3052_dilations_0, groups = var_3052_groups_0, pad = var_3052_pad_0, pad_type = var_3052_pad_type_0, strides = var_3052_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_3029_cast_fp16_0)[name = string("op_3052_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_79_cast_fp16 = mul(x = var_3046_cast_fp16, y = var_3052_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(827990272)))];
            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_3070_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3070_cast_fp16")];
            int32 var_3068 = const()[name = string("op_3068"), 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_3068, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3070_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(853156160)))];
            fp16 var_3080_to_fp16 = const()[name = string("op_3080_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_3080_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3091_split_sizes_0 = const()[name = string("op_3091_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3091_axis_0 = const()[name = string("op_3091_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3091_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3091_cast_fp16_1 = split(axis = var_3091_axis_0, split_sizes = var_3091_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3091_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(853164416)))];
            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_3091_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(861553088)))];
            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_3091_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            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_cast_fp16, x = var_3091_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_3148_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3148_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_3155_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3155_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3159_cast_fp16 = mul(x = x_81_cast_fp16, y = var_452_cast_fp16)[name = string("op_3159_cast_fp16")];
            tensor<int32, [2]> var_3160_split_sizes_0 = const()[name = string("op_3160_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3160_axis_0 = const()[name = string("op_3160_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_3160_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_3160_cast_fp16_1 = split(axis = var_3160_axis_0, split_sizes = var_3160_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3160_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_3162_cast_fp16 = mul(x = var_3160_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3162_cast_fp16")];
            int32 var_3164 = const()[name = string("op_3164"), val = int32(-2)];
            bool var_3165_interleave_0 = const()[name = string("op_3165_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_3165_cast_fp16 = concat(axis = var_3164, interleave = var_3165_interleave_0, values = (var_3162_cast_fp16, var_3160_cast_fp16_0))[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3166_cast_fp16 = mul(x = var_3165_cast_fp16, y = var_459_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_51_cast_fp16 = add(x = var_3159_cast_fp16, y = var_3166_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3172_cast_fp16 = mul(x = var_3148_cast_fp16, y = var_452_cast_fp16)[name = string("op_3172_cast_fp16")];
            tensor<int32, [2]> var_3173_split_sizes_0 = const()[name = string("op_3173_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3173_axis_0 = const()[name = string("op_3173_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_3173_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_3173_cast_fp16_1 = split(axis = var_3173_axis_0, split_sizes = var_3173_split_sizes_0, x = var_3148_cast_fp16)[name = string("op_3173_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_3175_cast_fp16 = mul(x = var_3173_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3175_cast_fp16")];
            int32 var_3177 = const()[name = string("op_3177"), val = int32(-2)];
            bool var_3178_interleave_0 = const()[name = string("op_3178_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_3178_cast_fp16 = concat(axis = var_3177, interleave = var_3178_interleave_0, values = (var_3175_cast_fp16, var_3173_cast_fp16_0))[name = string("op_3178_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3179_cast_fp16 = mul(x = var_3178_cast_fp16, y = var_459_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_85_cast_fp16 = add(x = var_3172_cast_fp16, y = var_3179_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_197")];
            tensor<fp16, [14, 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_126)[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_128_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_128 = read_state(input = key_cache)[name = string("coreml_update_state_128")];
            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_3155_cast_fp16)[name = string("transpose_196")];
            tensor<fp16, [14, 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_127)[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_129_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_129 = read_state(input = value_cache)[name = string("coreml_update_state_129")];
            tensor<int32, [4]> var_3249_begin_0 = const()[name = string("op_3249_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3249_end_0 = const()[name = string("op_3249_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3249_end_mask_0 = const()[name = string("op_3249_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3249_cast_fp16 = slice_by_index(begin = var_3249_begin_0, end = var_3249_end_0, end_mask = var_3249_end_mask_0, x = coreml_update_state_128)[name = string("op_3249_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3252_axis_0 = const()[name = string("op_3252_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_1 = split(axis = var_3252_axis_0, split_sizes = tile_16, x = var_3249_cast_fp16)[name = string("op_3252_cast_fp16")];
            tensor<int32, [4]> var_3259_begin_0 = const()[name = string("op_3259_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3259_end_0 = const()[name = string("op_3259_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3259_end_mask_0 = const()[name = string("op_3259_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3259_cast_fp16 = slice_by_index(begin = var_3259_begin_0, end = var_3259_end_0, end_mask = var_3259_end_mask_0, x = coreml_update_state_129)[name = string("op_3259_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3262_axis_0 = const()[name = string("op_3262_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_1 = split(axis = var_3262_axis_0, split_sizes = tile_17, x = var_3259_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<int32, [2]> var_3265_split_sizes_0 = const()[name = string("op_3265_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3265_axis_0 = const()[name = string("op_3265_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_3265_0, tensor<fp16, [1, 8, 128, 32]> var_3265_1 = split(axis = var_3265_axis_0, split_sizes = var_3265_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3265")];
            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_3252_cast_fp16_0, y = var_3265_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3268_to_fp16 = const()[name = string("op_3268_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_3268_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_3272 = const()[name = string("op_3272"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_135_cast_fp16 = softmax(axis = var_3272, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3278_transpose_x_1 = const()[name = string("op_3278_transpose_x_1"), val = bool(true)];
            bool var_3278_transpose_y_1 = const()[name = string("op_3278_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3278_cast_fp16 = matmul(transpose_x = var_3278_transpose_x_1, transpose_y = var_3278_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3262_cast_fp16_0)[name = string("op_3278_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_3252_cast_fp16_1, y = var_3265_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3280_to_fp16 = const()[name = string("op_3280_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_3280_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_3284 = const()[name = string("op_3284"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_143_cast_fp16 = softmax(axis = var_3284, x = attn_weights_141_cast_fp16)[name = string("attn_weights_143_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_143_cast_fp16, y = var_3262_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3292 = const()[name = string("op_3292"), 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_3292, interleave = attn_output_67_interleave_0, values = (var_3278_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3296_perm_0 = const()[name = string("op_3296_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_3296_cast_fp16 = transpose(perm = var_3296_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_195")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3296_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            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_cast_fp16, x = attn_output_71_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_3329_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3329_cast_fp16")];
            int32 var_3327 = const()[name = string("op_3327"), 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_3327, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3329_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(862601728)))];
            fp16 var_3339_to_fp16 = const()[name = string("op_3339_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_3339_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3350_split_sizes_0 = const()[name = string("op_3350_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3350_axis_0 = const()[name = string("op_3350_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3350_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3350_cast_fp16_1 = split(axis = var_3350_axis_0, split_sizes = var_3350_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3350_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_3367_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3367_cast_fp16")];
            tensor<int32, [2]> var_3373_strides_0 = const()[name = string("op_3373_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3373_pad_type_0 = const()[name = string("op_3373_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3373_pad_0 = const()[name = string("op_3373_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3373_dilations_0 = const()[name = string("op_3373_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3373_groups_0 = const()[name = string("op_3373_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_3373_cast_fp16 = conv(dilations = var_3373_dilations_0, groups = var_3373_groups_0, pad = var_3373_pad_0, pad_type = var_3373_pad_type_0, strides = var_3373_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("op_3373_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_89_cast_fp16 = mul(x = var_3367_cast_fp16, y = var_3373_cast_fp16)[name = string("x_89_cast_fp16")];
            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_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_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_3391_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3391_cast_fp16")];
            int32 var_3389 = const()[name = string("op_3389"), 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_3389, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3391_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_37_gamma_0_to_fp16 = const()[name = string("out_37_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862609984)))];
            fp16 var_3401_to_fp16 = const()[name = string("op_3401_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_3401_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3412_split_sizes_0 = const()[name = string("op_3412_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3412_axis_0 = const()[name = string("op_3412_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3412_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3412_cast_fp16_1 = split(axis = var_3412_axis_0, split_sizes = var_3412_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3412_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862618240)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(871006912)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3412_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_3469_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3469_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_3476_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3480_cast_fp16 = mul(x = x_91_cast_fp16, y = var_452_cast_fp16)[name = string("op_3480_cast_fp16")];
            tensor<int32, [2]> var_3481_split_sizes_0 = const()[name = string("op_3481_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3481_axis_0 = const()[name = string("op_3481_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_3481_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_3481_cast_fp16_1 = split(axis = var_3481_axis_0, split_sizes = var_3481_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3481_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_3483_cast_fp16 = mul(x = var_3481_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3483_cast_fp16")];
            int32 var_3485 = const()[name = string("op_3485"), val = int32(-2)];
            bool var_3486_interleave_0 = const()[name = string("op_3486_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_3486_cast_fp16 = concat(axis = var_3485, interleave = var_3486_interleave_0, values = (var_3483_cast_fp16, var_3481_cast_fp16_0))[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3487_cast_fp16 = mul(x = var_3486_cast_fp16, y = var_459_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_57_cast_fp16 = add(x = var_3480_cast_fp16, y = var_3487_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3493_cast_fp16 = mul(x = var_3469_cast_fp16, y = var_452_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [2]> var_3494_split_sizes_0 = const()[name = string("op_3494_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3494_axis_0 = const()[name = string("op_3494_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_3494_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_3494_cast_fp16_1 = split(axis = var_3494_axis_0, split_sizes = var_3494_split_sizes_0, x = var_3469_cast_fp16)[name = string("op_3494_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_3496_cast_fp16 = mul(x = var_3494_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3496_cast_fp16")];
            int32 var_3498 = const()[name = string("op_3498"), val = int32(-2)];
            bool var_3499_interleave_0 = const()[name = string("op_3499_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_3499_cast_fp16 = concat(axis = var_3498, interleave = var_3499_interleave_0, values = (var_3496_cast_fp16, var_3494_cast_fp16_0))[name = string("op_3499_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3500_cast_fp16 = mul(x = var_3499_cast_fp16, y = var_459_cast_fp16)[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_95_cast_fp16 = add(x = var_3493_cast_fp16, y = var_3500_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_194")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_128)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_130_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_130 = read_state(input = key_cache)[name = string("coreml_update_state_130")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3476_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_129)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_131_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_131 = read_state(input = value_cache)[name = string("coreml_update_state_131")];
            tensor<int32, [4]> var_3570_begin_0 = const()[name = string("op_3570_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3570_end_0 = const()[name = string("op_3570_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3570_end_mask_0 = const()[name = string("op_3570_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = coreml_update_state_130)[name = string("op_3570_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3573_axis_0 = const()[name = string("op_3573_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_1 = split(axis = var_3573_axis_0, split_sizes = tile_18, x = var_3570_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [4]> var_3580_begin_0 = const()[name = string("op_3580_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3580_end_0 = const()[name = string("op_3580_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3580_end_mask_0 = const()[name = string("op_3580_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3580_cast_fp16 = slice_by_index(begin = var_3580_begin_0, end = var_3580_end_0, end_mask = var_3580_end_mask_0, x = coreml_update_state_131)[name = string("op_3580_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3583_axis_0 = const()[name = string("op_3583_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_1 = split(axis = var_3583_axis_0, split_sizes = tile_19, x = var_3580_cast_fp16)[name = string("op_3583_cast_fp16")];
            tensor<int32, [2]> var_3586_split_sizes_0 = const()[name = string("op_3586_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3586_axis_0 = const()[name = string("op_3586_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_3586_0, tensor<fp16, [1, 8, 128, 32]> var_3586_1 = split(axis = var_3586_axis_0, split_sizes = var_3586_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3586")];
            bool attn_weights_145_transpose_x_0 = const()[name = string("attn_weights_145_transpose_x_0"), val = bool(false)];
            bool attn_weights_145_transpose_y_0 = const()[name = string("attn_weights_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_145_cast_fp16 = matmul(transpose_x = attn_weights_145_transpose_x_0, transpose_y = attn_weights_145_transpose_y_0, x = var_3573_cast_fp16_0, y = var_3586_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3589_to_fp16 = const()[name = string("op_3589_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_147_cast_fp16 = mul(x = attn_weights_145_cast_fp16, y = var_3589_to_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_149_cast_fp16 = add(x = attn_weights_147_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_149_cast_fp16")];
            int32 var_3593 = const()[name = string("op_3593"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_151_cast_fp16 = softmax(axis = var_3593, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3599_transpose_x_1 = const()[name = string("op_3599_transpose_x_1"), val = bool(true)];
            bool var_3599_transpose_y_1 = const()[name = string("op_3599_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3599_cast_fp16 = matmul(transpose_x = var_3599_transpose_x_1, transpose_y = var_3599_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3583_cast_fp16_0)[name = string("op_3599_cast_fp16")];
            bool attn_weights_153_transpose_x_0 = const()[name = string("attn_weights_153_transpose_x_0"), val = bool(false)];
            bool attn_weights_153_transpose_y_0 = const()[name = string("attn_weights_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_153_cast_fp16 = matmul(transpose_x = attn_weights_153_transpose_x_0, transpose_y = attn_weights_153_transpose_y_0, x = var_3573_cast_fp16_1, y = var_3586_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3601_to_fp16 = const()[name = string("op_3601_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_155_cast_fp16 = mul(x = attn_weights_153_cast_fp16, y = var_3601_to_fp16)[name = string("attn_weights_155_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_157_cast_fp16 = add(x = attn_weights_155_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_157_cast_fp16")];
            int32 var_3605 = const()[name = string("op_3605"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_159_cast_fp16 = softmax(axis = var_3605, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3583_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3613 = const()[name = string("op_3613"), val = int32(1)];
            bool attn_output_75_interleave_0 = const()[name = string("attn_output_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_75_cast_fp16 = concat(axis = var_3613, interleave = attn_output_75_interleave_0, values = (var_3599_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3617_perm_0 = const()[name = string("op_3617_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_3617_cast_fp16 = transpose(perm = var_3617_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_192")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3617_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_3650_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3650_cast_fp16")];
            int32 var_3648 = const()[name = string("op_3648"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_77_cast_fp16 = concat(axis = var_3648, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3650_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872055552)))];
            fp16 var_3660_to_fp16 = const()[name = string("op_3660_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3660_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3671_split_sizes_0 = const()[name = string("op_3671_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3671_axis_0 = const()[name = string("op_3671_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3671_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3671_cast_fp16_1 = split(axis = var_3671_axis_0, split_sizes = var_3671_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3671_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_3688_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3688_cast_fp16")];
            tensor<int32, [2]> var_3694_strides_0 = const()[name = string("op_3694_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3694_pad_type_0 = const()[name = string("op_3694_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3694_pad_0 = const()[name = string("op_3694_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3694_dilations_0 = const()[name = string("op_3694_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3694_groups_0 = const()[name = string("op_3694_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_3694_cast_fp16 = conv(dilations = var_3694_dilations_0, groups = var_3694_groups_0, pad = var_3694_pad_0, pad_type = var_3694_pad_type_0, strides = var_3694_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("op_3694_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_99_cast_fp16 = mul(x = var_3688_cast_fp16, y = var_3694_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_3712_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3712_cast_fp16")];
            int32 var_3710 = const()[name = string("op_3710"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_81_cast_fp16 = concat(axis = var_3710, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3712_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872063808)))];
            fp16 var_3722_to_fp16 = const()[name = string("op_3722_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3722_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3733_split_sizes_0 = const()[name = string("op_3733_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3733_axis_0 = const()[name = string("op_3733_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3733_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3733_cast_fp16_1 = split(axis = var_3733_axis_0, split_sizes = var_3733_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872072064)))];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(880460736)))];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_cast_fp16, x = var_3733_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_3790_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3790_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_3797_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3797_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3801_cast_fp16 = mul(x = x_101_cast_fp16, y = var_452_cast_fp16)[name = string("op_3801_cast_fp16")];
            tensor<int32, [2]> var_3802_split_sizes_0 = const()[name = string("op_3802_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3802_axis_0 = const()[name = string("op_3802_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_3802_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_3802_cast_fp16_1 = split(axis = var_3802_axis_0, split_sizes = var_3802_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3802_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_3804_cast_fp16 = mul(x = var_3802_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3804_cast_fp16")];
            int32 var_3806 = const()[name = string("op_3806"), val = int32(-2)];
            bool var_3807_interleave_0 = const()[name = string("op_3807_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_3807_cast_fp16 = concat(axis = var_3806, interleave = var_3807_interleave_0, values = (var_3804_cast_fp16, var_3802_cast_fp16_0))[name = string("op_3807_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_3808_cast_fp16 = mul(x = var_3807_cast_fp16, y = var_459_cast_fp16)[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_63_cast_fp16 = add(x = var_3801_cast_fp16, y = var_3808_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3814_cast_fp16 = mul(x = var_3790_cast_fp16, y = var_452_cast_fp16)[name = string("op_3814_cast_fp16")];
            tensor<int32, [2]> var_3815_split_sizes_0 = const()[name = string("op_3815_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3815_axis_0 = const()[name = string("op_3815_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_3815_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_3815_cast_fp16_1 = split(axis = var_3815_axis_0, split_sizes = var_3815_split_sizes_0, x = var_3790_cast_fp16)[name = string("op_3815_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_3817_cast_fp16 = mul(x = var_3815_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3817_cast_fp16")];
            int32 var_3819 = const()[name = string("op_3819"), val = int32(-2)];
            bool var_3820_interleave_0 = const()[name = string("op_3820_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_3820_cast_fp16 = concat(axis = var_3819, interleave = var_3820_interleave_0, values = (var_3817_cast_fp16, var_3815_cast_fp16_0))[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_3821_cast_fp16 = mul(x = var_3820_cast_fp16, y = var_459_cast_fp16)[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_105_cast_fp16 = add(x = var_3814_cast_fp16, y = var_3821_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_191")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_130)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_132_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_132 = read_state(input = key_cache)[name = string("coreml_update_state_132")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3797_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_131)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_133_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_133 = read_state(input = value_cache)[name = string("coreml_update_state_133")];
            tensor<int32, [4]> var_3891_begin_0 = const()[name = string("op_3891_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3891_end_0 = const()[name = string("op_3891_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3891_end_mask_0 = const()[name = string("op_3891_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3891_cast_fp16 = slice_by_index(begin = var_3891_begin_0, end = var_3891_end_0, end_mask = var_3891_end_mask_0, x = coreml_update_state_132)[name = string("op_3891_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3894_axis_0 = const()[name = string("op_3894_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_1 = split(axis = var_3894_axis_0, split_sizes = tile_20, x = var_3891_cast_fp16)[name = string("op_3894_cast_fp16")];
            tensor<int32, [4]> var_3901_begin_0 = const()[name = string("op_3901_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3901_end_0 = const()[name = string("op_3901_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3901_end_mask_0 = const()[name = string("op_3901_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3901_cast_fp16 = slice_by_index(begin = var_3901_begin_0, end = var_3901_end_0, end_mask = var_3901_end_mask_0, x = coreml_update_state_133)[name = string("op_3901_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3904_axis_0 = const()[name = string("op_3904_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_1 = split(axis = var_3904_axis_0, split_sizes = tile_21, x = var_3901_cast_fp16)[name = string("op_3904_cast_fp16")];
            tensor<int32, [2]> var_3907_split_sizes_0 = const()[name = string("op_3907_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3907_axis_0 = const()[name = string("op_3907_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_3907_0, tensor<fp16, [1, 8, 128, 32]> var_3907_1 = split(axis = var_3907_axis_0, split_sizes = var_3907_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3907")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3894_cast_fp16_0, y = var_3907_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3910_to_fp16 = const()[name = string("op_3910_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3910_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3914 = const()[name = string("op_3914"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_167_cast_fp16 = softmax(axis = var_3914, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3920_transpose_x_1 = const()[name = string("op_3920_transpose_x_1"), val = bool(true)];
            bool var_3920_transpose_y_1 = const()[name = string("op_3920_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3920_cast_fp16 = matmul(transpose_x = var_3920_transpose_x_1, transpose_y = var_3920_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3904_cast_fp16_0)[name = string("op_3920_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3894_cast_fp16_1, y = var_3907_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3922_to_fp16 = const()[name = string("op_3922_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3922_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3926 = const()[name = string("op_3926"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_175_cast_fp16 = softmax(axis = var_3926, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3904_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3934 = const()[name = string("op_3934"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_83_cast_fp16 = concat(axis = var_3934, interleave = attn_output_83_interleave_0, values = (var_3920_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3938_perm_0 = const()[name = string("op_3938_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_3938_cast_fp16 = transpose(perm = var_3938_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_189")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3938_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_3971_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3971_cast_fp16")];
            int32 var_3969 = const()[name = string("op_3969"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_85_cast_fp16 = concat(axis = var_3969, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3971_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881509376)))];
            fp16 var_3981_to_fp16 = const()[name = string("op_3981_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3981_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3992_split_sizes_0 = const()[name = string("op_3992_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3992_axis_0 = const()[name = string("op_3992_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_3992_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_3992_cast_fp16_1 = split(axis = var_3992_axis_0, split_sizes = var_3992_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3992_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_4009_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [2]> var_4015_strides_0 = const()[name = string("op_4015_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4015_pad_type_0 = const()[name = string("op_4015_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4015_pad_0 = const()[name = string("op_4015_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4015_dilations_0 = const()[name = string("op_4015_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4015_groups_0 = const()[name = string("op_4015_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_4015_cast_fp16 = conv(dilations = var_4015_dilations_0, groups = var_4015_groups_0, pad = var_4015_pad_0, pad_type = var_4015_pad_type_0, strides = var_4015_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("op_4015_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_109_cast_fp16 = mul(x = var_4009_cast_fp16, y = var_4015_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_4033_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4033_cast_fp16")];
            int32 var_4031 = const()[name = string("op_4031"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_89_cast_fp16 = concat(axis = var_4031, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4033_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881517632)))];
            fp16 var_4043_to_fp16 = const()[name = string("op_4043_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4043_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4054_split_sizes_0 = const()[name = string("op_4054_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4054_axis_0 = const()[name = string("op_4054_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_4054_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_4054_cast_fp16_1 = split(axis = var_4054_axis_0, split_sizes = var_4054_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_4111_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4111_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_4118_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4118_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_4122_cast_fp16 = mul(x = x_111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4122_cast_fp16")];
            tensor<int32, [2]> var_4123_split_sizes_0 = const()[name = string("op_4123_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4123_axis_0 = const()[name = string("op_4123_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_4123_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_4123_cast_fp16_1 = split(axis = var_4123_axis_0, split_sizes = var_4123_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4123_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_4125_cast_fp16 = mul(x = var_4123_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4125_cast_fp16")];
            int32 var_4127 = const()[name = string("op_4127"), val = int32(-2)];
            bool var_4128_interleave_0 = const()[name = string("op_4128_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_4128_cast_fp16 = concat(axis = var_4127, interleave = var_4128_interleave_0, values = (var_4125_cast_fp16, var_4123_cast_fp16_0))[name = string("op_4128_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_4129_cast_fp16 = mul(x = var_4128_cast_fp16, y = var_459_cast_fp16)[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_69_cast_fp16 = add(x = var_4122_cast_fp16, y = var_4129_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_4135_cast_fp16 = mul(x = var_4111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4135_cast_fp16")];
            tensor<int32, [2]> var_4136_split_sizes_0 = const()[name = string("op_4136_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4136_axis_0 = const()[name = string("op_4136_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_4136_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_4136_cast_fp16_1 = split(axis = var_4136_axis_0, split_sizes = var_4136_split_sizes_0, x = var_4111_cast_fp16)[name = string("op_4136_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_4138_cast_fp16 = mul(x = var_4136_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4138_cast_fp16")];
            int32 var_4140 = const()[name = string("op_4140"), val = int32(-2)];
            bool var_4141_interleave_0 = const()[name = string("op_4141_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_4141_cast_fp16 = concat(axis = var_4140, interleave = var_4141_interleave_0, values = (var_4138_cast_fp16, var_4136_cast_fp16_0))[name = string("op_4141_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_4142_cast_fp16 = mul(x = var_4141_cast_fp16, y = var_459_cast_fp16)[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_115_cast_fp16 = add(x = var_4135_cast_fp16, y = var_4142_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_132)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_134_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_134 = read_state(input = key_cache)[name = string("coreml_update_state_134")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4118_cast_fp16)[name = string("transpose_187")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_133)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_135_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_135 = read_state(input = value_cache)[name = string("coreml_update_state_135")];
            tensor<int32, [4]> var_4212_begin_0 = const()[name = string("op_4212_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4212_end_0 = const()[name = string("op_4212_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4212_end_mask_0 = const()[name = string("op_4212_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4212_cast_fp16 = slice_by_index(begin = var_4212_begin_0, end = var_4212_end_0, end_mask = var_4212_end_mask_0, x = coreml_update_state_134)[name = string("op_4212_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4215_axis_0 = const()[name = string("op_4215_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_1 = split(axis = var_4215_axis_0, split_sizes = tile_22, x = var_4212_cast_fp16)[name = string("op_4215_cast_fp16")];
            tensor<int32, [4]> var_4222_begin_0 = const()[name = string("op_4222_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4222_end_0 = const()[name = string("op_4222_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4222_end_mask_0 = const()[name = string("op_4222_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4222_cast_fp16 = slice_by_index(begin = var_4222_begin_0, end = var_4222_end_0, end_mask = var_4222_end_mask_0, x = coreml_update_state_135)[name = string("op_4222_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4225_axis_0 = const()[name = string("op_4225_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_1 = split(axis = var_4225_axis_0, split_sizes = tile_23, x = var_4222_cast_fp16)[name = string("op_4225_cast_fp16")];
            tensor<int32, [2]> var_4228_split_sizes_0 = const()[name = string("op_4228_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4228_axis_0 = const()[name = string("op_4228_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_4228_0, tensor<fp16, [1, 8, 128, 32]> var_4228_1 = split(axis = var_4228_axis_0, split_sizes = var_4228_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4228")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4215_cast_fp16_0, y = var_4228_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4231_to_fp16 = const()[name = string("op_4231_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4231_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4235 = const()[name = string("op_4235"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_183_cast_fp16 = softmax(axis = var_4235, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4241_transpose_x_1 = const()[name = string("op_4241_transpose_x_1"), val = bool(true)];
            bool var_4241_transpose_y_1 = const()[name = string("op_4241_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4241_cast_fp16 = matmul(transpose_x = var_4241_transpose_x_1, transpose_y = var_4241_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4225_cast_fp16_0)[name = string("op_4241_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4215_cast_fp16_1, y = var_4228_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4243_to_fp16 = const()[name = string("op_4243_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4243_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4247 = const()[name = string("op_4247"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_191_cast_fp16 = softmax(axis = var_4247, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4225_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4255 = const()[name = string("op_4255"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_91_cast_fp16 = concat(axis = var_4255, interleave = attn_output_91_interleave_0, values = (var_4241_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4259_perm_0 = const()[name = string("op_4259_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_4259_cast_fp16 = transpose(perm = var_4259_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_186")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4259_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_4292_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4292_cast_fp16")];
            int32 var_4290 = const()[name = string("op_4290"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_93_cast_fp16 = concat(axis = var_4290, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4292_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881525888)))];
            fp16 var_4302_to_fp16 = const()[name = string("op_4302_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4302_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4313_split_sizes_0 = const()[name = string("op_4313_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4313_axis_0 = const()[name = string("op_4313_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_4313_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_4313_cast_fp16_1 = split(axis = var_4313_axis_0, split_sizes = var_4313_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4313_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_4330_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4330_cast_fp16")];
            tensor<int32, [2]> var_4336_strides_0 = const()[name = string("op_4336_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4336_pad_type_0 = const()[name = string("op_4336_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4336_pad_0 = const()[name = string("op_4336_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4336_dilations_0 = const()[name = string("op_4336_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4336_groups_0 = const()[name = string("op_4336_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_4336_cast_fp16 = conv(dilations = var_4336_dilations_0, groups = var_4336_groups_0, pad = var_4336_pad_0, pad_type = var_4336_pad_type_0, strides = var_4336_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("op_4336_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_119_cast_fp16 = mul(x = var_4330_cast_fp16, y = var_4336_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_cast_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_4354_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4354_cast_fp16")];
            int32 var_4352 = const()[name = string("op_4352"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_97_cast_fp16 = concat(axis = var_4352, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4354_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881534144)))];
            fp16 var_4364_to_fp16 = const()[name = string("op_4364_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4364_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4375_split_sizes_0 = const()[name = string("op_4375_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4375_axis_0 = const()[name = string("op_4375_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_4375_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_4375_cast_fp16_1 = split(axis = var_4375_axis_0, split_sizes = var_4375_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4375_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_4432_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4432_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_4439_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4439_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_4443_cast_fp16 = mul(x = x_121_cast_fp16, y = var_452_cast_fp16)[name = string("op_4443_cast_fp16")];
            tensor<int32, [2]> var_4444_split_sizes_0 = const()[name = string("op_4444_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4444_axis_0 = const()[name = string("op_4444_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_4444_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_4444_cast_fp16_1 = split(axis = var_4444_axis_0, split_sizes = var_4444_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4444_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_4446_cast_fp16 = mul(x = var_4444_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4446_cast_fp16")];
            int32 var_4448 = const()[name = string("op_4448"), val = int32(-2)];
            bool var_4449_interleave_0 = const()[name = string("op_4449_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_4449_cast_fp16 = concat(axis = var_4448, interleave = var_4449_interleave_0, values = (var_4446_cast_fp16, var_4444_cast_fp16_0))[name = string("op_4449_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_4450_cast_fp16 = mul(x = var_4449_cast_fp16, y = var_459_cast_fp16)[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_75_cast_fp16 = add(x = var_4443_cast_fp16, y = var_4450_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_4456_cast_fp16 = mul(x = var_4432_cast_fp16, y = var_452_cast_fp16)[name = string("op_4456_cast_fp16")];
            tensor<int32, [2]> var_4457_split_sizes_0 = const()[name = string("op_4457_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4457_axis_0 = const()[name = string("op_4457_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_4457_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_4457_cast_fp16_1 = split(axis = var_4457_axis_0, split_sizes = var_4457_split_sizes_0, x = var_4432_cast_fp16)[name = string("op_4457_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_4459_cast_fp16 = mul(x = var_4457_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4459_cast_fp16")];
            int32 var_4461 = const()[name = string("op_4461"), val = int32(-2)];
            bool var_4462_interleave_0 = const()[name = string("op_4462_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_4462_cast_fp16 = concat(axis = var_4461, interleave = var_4462_interleave_0, values = (var_4459_cast_fp16, var_4457_cast_fp16_0))[name = string("op_4462_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_4463_cast_fp16 = mul(x = var_4462_cast_fp16, y = var_459_cast_fp16)[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_125_cast_fp16 = add(x = var_4456_cast_fp16, y = var_4463_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_185")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_134)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_136_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_136 = read_state(input = key_cache)[name = string("coreml_update_state_136")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4439_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_135)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_137_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_137 = read_state(input = value_cache)[name = string("coreml_update_state_137")];
            tensor<int32, [4]> var_4533_begin_0 = const()[name = string("op_4533_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4533_end_0 = const()[name = string("op_4533_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4533_end_mask_0 = const()[name = string("op_4533_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4533_cast_fp16 = slice_by_index(begin = var_4533_begin_0, end = var_4533_end_0, end_mask = var_4533_end_mask_0, x = coreml_update_state_136)[name = string("op_4533_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4536_axis_0 = const()[name = string("op_4536_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_1 = split(axis = var_4536_axis_0, split_sizes = tile_24, x = var_4533_cast_fp16)[name = string("op_4536_cast_fp16")];
            tensor<int32, [4]> var_4543_begin_0 = const()[name = string("op_4543_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4543_end_0 = const()[name = string("op_4543_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4543_end_mask_0 = const()[name = string("op_4543_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4543_cast_fp16 = slice_by_index(begin = var_4543_begin_0, end = var_4543_end_0, end_mask = var_4543_end_mask_0, x = coreml_update_state_137)[name = string("op_4543_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4546_axis_0 = const()[name = string("op_4546_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_1 = split(axis = var_4546_axis_0, split_sizes = tile_25, x = var_4543_cast_fp16)[name = string("op_4546_cast_fp16")];
            tensor<int32, [2]> var_4549_split_sizes_0 = const()[name = string("op_4549_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4549_axis_0 = const()[name = string("op_4549_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_4549_0, tensor<fp16, [1, 8, 128, 32]> var_4549_1 = split(axis = var_4549_axis_0, split_sizes = var_4549_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4549")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4536_cast_fp16_0, y = var_4549_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4552_to_fp16 = const()[name = string("op_4552_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4552_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4556 = const()[name = string("op_4556"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_199_cast_fp16 = softmax(axis = var_4556, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4562_transpose_x_1 = const()[name = string("op_4562_transpose_x_1"), val = bool(true)];
            bool var_4562_transpose_y_1 = const()[name = string("op_4562_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4562_cast_fp16 = matmul(transpose_x = var_4562_transpose_x_1, transpose_y = var_4562_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4546_cast_fp16_0)[name = string("op_4562_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4536_cast_fp16_1, y = var_4549_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4564_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4568 = const()[name = string("op_4568"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_207_cast_fp16 = softmax(axis = var_4568, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4546_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4576 = const()[name = string("op_4576"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_99_cast_fp16 = concat(axis = var_4576, interleave = attn_output_99_interleave_0, values = (var_4562_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4580_perm_0 = const()[name = string("op_4580_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_4580_cast_fp16 = transpose(perm = var_4580_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_183")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4580_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_4613_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4613_cast_fp16")];
            int32 var_4611 = const()[name = string("op_4611"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_101_cast_fp16 = concat(axis = var_4611, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4613_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881542400)))];
            fp16 var_4623_to_fp16 = const()[name = string("op_4623_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4623_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4634_split_sizes_0 = const()[name = string("op_4634_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4634_axis_0 = const()[name = string("op_4634_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_4634_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_4634_cast_fp16_1 = split(axis = var_4634_axis_0, split_sizes = var_4634_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4634_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_4651_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4651_cast_fp16")];
            tensor<int32, [2]> var_4657_strides_0 = const()[name = string("op_4657_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4657_pad_type_0 = const()[name = string("op_4657_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4657_pad_0 = const()[name = string("op_4657_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4657_dilations_0 = const()[name = string("op_4657_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4657_groups_0 = const()[name = string("op_4657_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_4657_cast_fp16 = conv(dilations = var_4657_dilations_0, groups = var_4657_groups_0, pad = var_4657_pad_0, pad_type = var_4657_pad_type_0, strides = var_4657_strides_0, weight = layers_12_mlp_up_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("op_4657_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_129_cast_fp16 = mul(x = var_4651_cast_fp16, y = var_4657_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_cast_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_4675_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4675_cast_fp16")];
            int32 var_4673 = const()[name = string("op_4673"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_105_cast_fp16 = concat(axis = var_4673, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4675_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881550656)))];
            fp16 var_4685_to_fp16 = const()[name = string("op_4685_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4685_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4696_split_sizes_0 = const()[name = string("op_4696_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4696_axis_0 = const()[name = string("op_4696_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_4696_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_4696_cast_fp16_1 = split(axis = var_4696_axis_0, split_sizes = var_4696_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4696_cast_fp16")];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 32]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 32]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_4753_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4753_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 32]> var_4760_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4760_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_4764_cast_fp16 = mul(x = x_131_cast_fp16, y = var_452_cast_fp16)[name = string("op_4764_cast_fp16")];
            tensor<int32, [2]> var_4765_split_sizes_0 = const()[name = string("op_4765_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4765_axis_0 = const()[name = string("op_4765_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 32]> var_4765_cast_fp16_0, tensor<fp16, [1, 16, 64, 32]> var_4765_cast_fp16_1 = split(axis = var_4765_axis_0, split_sizes = var_4765_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4765_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 32]> var_4767_cast_fp16 = mul(x = var_4765_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4767_cast_fp16")];
            int32 var_4769 = const()[name = string("op_4769"), val = int32(-2)];
            bool var_4770_interleave_0 = const()[name = string("op_4770_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 32]> var_4770_cast_fp16 = concat(axis = var_4769, interleave = var_4770_interleave_0, values = (var_4767_cast_fp16, var_4765_cast_fp16_0))[name = string("op_4770_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> var_4771_cast_fp16 = mul(x = var_4770_cast_fp16, y = var_459_cast_fp16)[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 32]> query_states_81_cast_fp16 = add(x = var_4764_cast_fp16, y = var_4771_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_4777_cast_fp16 = mul(x = var_4753_cast_fp16, y = var_452_cast_fp16)[name = string("op_4777_cast_fp16")];
            tensor<int32, [2]> var_4778_split_sizes_0 = const()[name = string("op_4778_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4778_axis_0 = const()[name = string("op_4778_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 32]> var_4778_cast_fp16_0, tensor<fp16, [1, 2, 64, 32]> var_4778_cast_fp16_1 = split(axis = var_4778_axis_0, split_sizes = var_4778_split_sizes_0, x = var_4753_cast_fp16)[name = string("op_4778_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_4780_cast_fp16 = mul(x = var_4778_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4780_cast_fp16")];
            int32 var_4782 = const()[name = string("op_4782"), val = int32(-2)];
            bool var_4783_interleave_0 = const()[name = string("op_4783_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 32]> var_4783_cast_fp16 = concat(axis = var_4782, interleave = var_4783_interleave_0, values = (var_4780_cast_fp16, var_4778_cast_fp16_0))[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> var_4784_cast_fp16 = mul(x = var_4783_cast_fp16, y = var_459_cast_fp16)[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 32]> key_states_135_cast_fp16 = add(x = var_4777_cast_fp16, y = var_4784_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_182")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_136)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_138_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_138 = read_state(input = key_cache)[name = string("coreml_update_state_138")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 32, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4760_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_137)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_139_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_139 = read_state(input = value_cache)[name = string("coreml_update_state_139")];
            tensor<int32, [4]> var_4854_begin_0 = const()[name = string("op_4854_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4854_end_0 = const()[name = string("op_4854_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4854_end_mask_0 = const()[name = string("op_4854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4854_cast_fp16 = slice_by_index(begin = var_4854_begin_0, end = var_4854_end_0, end_mask = var_4854_end_mask_0, x = coreml_update_state_138)[name = string("op_4854_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4857_axis_0 = const()[name = string("op_4857_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_1 = split(axis = var_4857_axis_0, split_sizes = tile_26, x = var_4854_cast_fp16)[name = string("op_4857_cast_fp16")];
            tensor<int32, [4]> var_4864_begin_0 = const()[name = string("op_4864_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4864_end_0 = const()[name = string("op_4864_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4864_end_mask_0 = const()[name = string("op_4864_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4864_cast_fp16 = slice_by_index(begin = var_4864_begin_0, end = var_4864_end_0, end_mask = var_4864_end_mask_0, x = coreml_update_state_139)[name = string("op_4864_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4867_axis_0 = const()[name = string("op_4867_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_1 = split(axis = var_4867_axis_0, split_sizes = tile_27, x = var_4864_cast_fp16)[name = string("op_4867_cast_fp16")];
            tensor<int32, [2]> var_4870_split_sizes_0 = const()[name = string("op_4870_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4870_axis_0 = const()[name = string("op_4870_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 32]> var_4870_0, tensor<fp16, [1, 8, 128, 32]> var_4870_1 = split(axis = var_4870_axis_0, split_sizes = var_4870_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4870")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4857_cast_fp16_0, y = var_4870_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4873_to_fp16 = const()[name = string("op_4873_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4873_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4877 = const()[name = string("op_4877"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_215_cast_fp16 = softmax(axis = var_4877, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4883_transpose_x_1 = const()[name = string("op_4883_transpose_x_1"), val = bool(true)];
            bool var_4883_transpose_y_1 = const()[name = string("op_4883_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4883_cast_fp16 = matmul(transpose_x = var_4883_transpose_x_1, transpose_y = var_4883_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4867_cast_fp16_0)[name = string("op_4883_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4857_cast_fp16_1, y = var_4870_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4885_to_fp16 = const()[name = string("op_4885_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4885_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4889 = const()[name = string("op_4889"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 32]> attn_weights_cast_fp16 = softmax(axis = var_4889, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4867_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4897 = const()[name = string("op_4897"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32, 128]> attn_output_107_cast_fp16 = concat(axis = var_4897, interleave = attn_output_107_interleave_0, values = (var_4883_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4901_perm_0 = const()[name = string("op_4901_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 32]> var_4901_cast_fp16 = transpose(perm = var_4901_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 2048, 1, 32]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4901_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 32]> var_4934_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4934_cast_fp16")];
            int32 var_4932 = const()[name = string("op_4932"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 32]> doubled_109_cast_fp16 = concat(axis = var_4932, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4934_cast_fp16))[name = string("doubled_109_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(881558912)))];
            fp16 var_4944_to_fp16 = const()[name = string("op_4944_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 32]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_4944_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_4955_split_sizes_0 = const()[name = string("op_4955_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4955_axis_0 = const()[name = string("op_4955_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> var_4955_cast_fp16_0, tensor<fp16, [1, 2048, 1, 32]> var_4955_cast_fp16_1 = split(axis = var_4955_axis_0, split_sizes = var_4955_split_sizes_0, x = out_cast_fp16)[name = string("op_4955_cast_fp16")];
            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_13_mlp_gate_proj_weight_cast_fp16, x = var_4955_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> var_4972_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4972_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881567168)))];
            tensor<int32, [2]> var_4978_strides_0 = const()[name = string("op_4978_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4978_pad_type_0 = const()[name = string("op_4978_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4978_pad_0 = const()[name = string("op_4978_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4978_dilations_0 = const()[name = string("op_4978_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4978_groups_0 = const()[name = string("op_4978_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 32]> var_4978_cast_fp16 = conv(dilations = var_4978_dilations_0, groups = var_4978_groups_0, pad = var_4978_pad_0, pad_type = var_4978_pad_type_0, strides = var_4978_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4955_cast_fp16_0)[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 32]> x_cast_fp16 = mul(x = var_4972_cast_fp16, y = var_4978_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_13_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 32]> hidden_states = add(x = hidden_states_135_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_4987_cast_fp16")];
        } -> (hidden_states);
    func length_64<ios18>(tensor<fp32, [1, 2048, 1, 64]> inputs_embeds, state<tensor<fp16, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [64]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_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(64))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524992))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524416))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_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(525312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13120640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13108288))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_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(13126848))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651776))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651200))))[name = string("layers_2_self_attn_v_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(13652096))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26247424))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26235072))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_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(26253632))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26778560))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26777984))))[name = string("layers_3_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_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(26778880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30977408))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30973248))))[name = string("layers_3_self_attn_o_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(30979520))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43566656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43562496))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_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(43568768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093120))))[name = string("layers_4_self_attn_v_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(44094016))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48292544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48288384))))[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(48294656))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60889984))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60877632))))[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(60896192))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73491520))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73479168))))[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(73497728))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86084864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86080704))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_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(86086976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611328))))[name = string("layers_5_self_attn_v_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(86612224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90810752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90806592))))[name = string("layers_5_self_attn_o_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(90812864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103395840))))[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(103414400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997376))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_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(116003648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528000))))[name = string("layers_6_self_attn_v_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(116528896))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120727424))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120723264))))[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(120729536))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133324864))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133312512))))[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(133331072))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145926400))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145914048))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_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(145932608))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158519744))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158515584))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_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(158521856))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046784))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046208))))[name = string("layers_7_self_attn_v_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(159047104))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163245632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241472))))[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(163247744))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175843072))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175830720))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_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(175849280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176374208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176373632))))[name = string("layers_8_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_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(176374528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180573056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180568896))))[name = string("layers_8_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(180575168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193170496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193158144))))[name = string("layers_8_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(193176704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205772032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205759680))))[name = string("layers_8_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_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(205778240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218365376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218361216))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_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(218367488))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218892416))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218891840))))[name = string("layers_9_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_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(218892736))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223091264))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223087104))))[name = string("layers_9_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(223093376))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235688704))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235676352))))[name = string("layers_9_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(235694912))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248290240))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248277888))))[name = string("layers_9_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_9_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(248296448))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260883584))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260879424))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_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(260885696))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410624))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410048))))[name = string("layers_10_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_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(261410944))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265609472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265605312))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(265611584))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278206912))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278194560))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(278213120))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290808448))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290796096))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_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(290814656))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303401792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303397632))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(303403904))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307602432))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307598272))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_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(307604544))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308129472))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308128896))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_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(308129792))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654720))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654144))))[name = string("layers_11_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(308655040))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312853568))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312849408))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(312855680))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325451008))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325438656))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(325457216))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338052544))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040192))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_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(338058752))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350645888))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350641728))))[name = string("layers_11_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(350648000))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354846528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354842368))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_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(354848640))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355373568))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355372992))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_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(355373888))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898816))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898240))))[name = string("layers_12_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(355899136))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360097664))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360093504))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(360099776))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372695104))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372682752))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(372701312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385296640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385284288))))[name = string("layers_12_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_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(385302848))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397889984))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397885824))))[name = string("layers_12_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(397892096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402090624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402086464))))[name = string("layers_13_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_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(402092736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617088))))[name = string("layers_13_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_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(402617984))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142912))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142336))))[name = string("layers_13_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(403143232))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407341760))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407337600))))[name = string("layers_13_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_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(407343872))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419939200))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419926848))))[name = string("layers_13_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_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(419945408))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432532544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432528384))))[name = string("layers_13_mlp_down_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_424 = const()[name = string("op_424"), val = int32(0)];
            bool var_426_exclusive_0 = const()[name = string("op_426_exclusive_0"), val = bool(false)];
            bool var_426_reverse_0 = const()[name = string("op_426_reverse_0"), val = bool(false)];
            tensor<fp16, [64]> var_426_cast_fp16 = cumsum(axis = var_424, exclusive = var_426_exclusive_0, reverse = var_426_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_426_cast_fp16")];
            fp16 var_428_promoted_to_fp16 = const()[name = string("op_428_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [64]> position_offsets_cast_fp16 = sub(x = var_426_cast_fp16, y = var_428_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_431_axes_0 = const()[name = string("op_431_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 64]> var_431_cast_fp16 = expand_dims(axes = var_431_axes_0, x = position_offsets_cast_fp16)[name = string("op_431_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_431_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(432534656)))];
            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(440923328)))];
            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_450_perm_0 = const()[name = string("op_450_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_452_axes_0 = const()[name = string("op_452_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 64]> var_450_cast_fp16 = transpose(perm = var_450_perm_0, x = cos_1_cast_fp16)[name = string("transpose_269")];
            tensor<fp16, [1, 1, 128, 64]> var_452_cast_fp16 = expand_dims(axes = var_452_axes_0, x = var_450_cast_fp16)[name = string("op_452_cast_fp16")];
            tensor<int32, [3]> var_457_perm_0 = const()[name = string("op_457_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_459_axes_0 = const()[name = string("op_459_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 64]> var_457_cast_fp16 = transpose(perm = var_457_perm_0, x = sin_1_cast_fp16)[name = string("transpose_268")];
            tensor<fp16, [1, 1, 128, 64]> var_459_cast_fp16 = expand_dims(axes = var_459_axes_0, x = var_457_cast_fp16)[name = string("op_459_cast_fp16")];
            tensor<int32, [1]> var_478_axes_0 = const()[name = string("op_478_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 64, 1]> var_478 = expand_dims(axes = var_478_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_478")];
            tensor<int32, [1, 1, 2048]> var_471 = const()[name = string("op_471"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449312000)))];
            tensor<bool, [1, 64, 2048]> var_479 = greater(x = var_471, y = var_478)[name = string("op_479")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_486_axes_0 = const()[name = string("op_486_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 64, 2048]> var_479_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_479)[name = string("cast_21")];
            tensor<fp16, [1, 1, 64, 2048]> var_486_cast_fp16 = expand_dims(axes = var_486_axes_0, x = var_479_to_fp16)[name = string("op_486_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_490_promoted_to_fp16 = const()[name = string("op_490_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 64]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_486_cast_fp16)[name = string("transpose_267")];
            tensor<bool, [1, 1, 2048, 64]> var_491_cast_fp16 = equal(x = mask_cast_fp16, y = var_490_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            fp16 var_492_to_fp16 = const()[name = string("op_492_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 64]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_492_to_fp16, cond = var_491_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_502_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_502_cast_fp16")];
            int32 var_500 = const()[name = string("op_500"), 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_500, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_502_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(449320256)))];
            fp16 var_512_to_fp16 = const()[name = string("op_512_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_512_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_523_split_sizes_0 = const()[name = string("op_523_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_523_axis_0 = const()[name = string("op_523_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_523_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_523_cast_fp16_1 = split(axis = var_523_axis_0, split_sizes = var_523_split_sizes_0, x = out_1_cast_fp16)[name = string("op_523_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449328512)))];
            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_to_fp16, x = var_523_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457717184)))];
            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_to_fp16, x = var_523_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(458765824)))];
            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_523_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_580_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_580_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_587_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_587_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_591_cast_fp16 = mul(x = x_1_cast_fp16, y = var_452_cast_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [2]> var_592_split_sizes_0 = const()[name = string("op_592_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_592_axis_0 = const()[name = string("op_592_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_592_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_592_cast_fp16_1 = split(axis = var_592_axis_0, split_sizes = var_592_split_sizes_0, x = x_1_cast_fp16)[name = string("op_592_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_594_cast_fp16 = mul(x = var_592_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_594_cast_fp16")];
            int32 var_596 = const()[name = string("op_596"), val = int32(-2)];
            bool var_597_interleave_0 = const()[name = string("op_597_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_597_cast_fp16 = concat(axis = var_596, interleave = var_597_interleave_0, values = (var_594_cast_fp16, var_592_cast_fp16_0))[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_598_cast_fp16 = mul(x = var_597_cast_fp16, y = var_459_cast_fp16)[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_3_cast_fp16 = add(x = var_591_cast_fp16, y = var_598_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_604_cast_fp16 = mul(x = var_580_cast_fp16, y = var_452_cast_fp16)[name = string("op_604_cast_fp16")];
            tensor<int32, [2]> var_605_split_sizes_0 = const()[name = string("op_605_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_605_axis_0 = const()[name = string("op_605_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_605_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_605_cast_fp16_1 = split(axis = var_605_axis_0, split_sizes = var_605_split_sizes_0, x = var_580_cast_fp16)[name = string("op_605_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_607_cast_fp16 = mul(x = var_605_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_607_cast_fp16")];
            int32 var_609 = const()[name = string("op_609"), val = int32(-2)];
            bool var_610_interleave_0 = const()[name = string("op_610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_610_cast_fp16 = concat(axis = var_609, interleave = var_610_interleave_0, values = (var_607_cast_fp16, var_605_cast_fp16_0))[name = string("op_610_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_611_cast_fp16 = mul(x = var_610_cast_fp16, y = var_459_cast_fp16)[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_5_cast_fp16 = add(x = var_604_cast_fp16, y = var_611_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 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_266")];
            tensor<fp16, [14, 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_140_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_140 = read_state(input = key_cache)[name = string("coreml_update_state_140")];
            tensor<fp16, [14, 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_587_cast_fp16)[name = string("transpose_265")];
            tensor<fp16, [14, 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_141_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_141 = read_state(input = value_cache)[name = string("coreml_update_state_141")];
            tensor<int32, [4]> var_681_begin_0 = const()[name = string("op_681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_681_end_0 = const()[name = string("op_681_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_681_end_mask_0 = const()[name = string("op_681_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_681_cast_fp16 = slice_by_index(begin = var_681_begin_0, end = var_681_end_0, end_mask = var_681_end_mask_0, x = coreml_update_state_140)[name = string("op_681_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_684_axis_0 = const()[name = string("op_684_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_1 = split(axis = var_684_axis_0, split_sizes = tile_0, x = var_681_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [4]> var_691_begin_0 = const()[name = string("op_691_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_691_end_0 = const()[name = string("op_691_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_691_end_mask_0 = const()[name = string("op_691_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = coreml_update_state_141)[name = string("op_691_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_694_axis_0 = const()[name = string("op_694_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_1 = split(axis = var_694_axis_0, split_sizes = tile_1, x = var_691_cast_fp16)[name = string("op_694_cast_fp16")];
            tensor<int32, [2]> var_697_split_sizes_0 = const()[name = string("op_697_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_697_axis_0 = const()[name = string("op_697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_697_0, tensor<fp16, [1, 8, 128, 64]> var_697_1 = split(axis = var_697_axis_0, split_sizes = var_697_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_697")];
            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_684_cast_fp16_0, y = var_697_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_700_to_fp16 = const()[name = string("op_700_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_700_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_704 = const()[name = string("op_704"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_7_cast_fp16 = softmax(axis = var_704, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_710_transpose_x_1 = const()[name = string("op_710_transpose_x_1"), val = bool(true)];
            bool var_710_transpose_y_1 = const()[name = string("op_710_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_710_cast_fp16 = matmul(transpose_x = var_710_transpose_x_1, transpose_y = var_710_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_694_cast_fp16_0)[name = string("op_710_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_684_cast_fp16_1, y = var_697_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_712_to_fp16 = const()[name = string("op_712_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_712_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_716 = const()[name = string("op_716"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_15_cast_fp16 = softmax(axis = var_716, 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_694_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_724 = const()[name = string("op_724"), 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_724, interleave = attn_output_3_interleave_0, values = (var_710_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_728_perm_0 = const()[name = string("op_728_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_728_cast_fp16 = transpose(perm = var_728_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_264")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_728_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(459814464)))];
            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_761_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_761_cast_fp16")];
            int32 var_759 = const()[name = string("op_759"), 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_759, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_761_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(468203136)))];
            fp16 var_771_to_fp16 = const()[name = string("op_771_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_771_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_782_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = out_3_cast_fp16)[name = string("op_782_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468211392)))];
            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_to_fp16, x = var_782_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_799_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_799_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493377280)))];
            tensor<int32, [2]> var_805_strides_0 = const()[name = string("op_805_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_805_pad_type_0 = const()[name = string("op_805_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_805_pad_0 = const()[name = string("op_805_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_805_dilations_0 = const()[name = string("op_805_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_805_groups_0 = const()[name = string("op_805_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_805_cast_fp16 = conv(dilations = var_805_dilations_0, groups = var_805_groups_0, pad = var_805_pad_0, pad_type = var_805_pad_type_0, strides = var_805_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_782_cast_fp16_0)[name = string("op_805_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_9_cast_fp16 = mul(x = var_799_cast_fp16, y = var_805_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518543168)))];
            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_to_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_823_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_823_cast_fp16")];
            int32 var_821 = const()[name = string("op_821"), 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_821, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_823_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(543709056)))];
            fp16 var_833_to_fp16 = const()[name = string("op_833_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_833_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_844_split_sizes_0 = const()[name = string("op_844_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_844_axis_0 = const()[name = string("op_844_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_844_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_844_cast_fp16_1 = split(axis = var_844_axis_0, split_sizes = var_844_split_sizes_0, x = out_5_cast_fp16)[name = string("op_844_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543717312)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552105984)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            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_cast_fp16, x = var_844_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_901_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_901_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_908_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_908_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_912_cast_fp16 = mul(x = x_11_cast_fp16, y = var_452_cast_fp16)[name = string("op_912_cast_fp16")];
            tensor<int32, [2]> var_913_split_sizes_0 = const()[name = string("op_913_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_913_axis_0 = const()[name = string("op_913_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_913_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_913_cast_fp16_1 = split(axis = var_913_axis_0, split_sizes = var_913_split_sizes_0, x = x_11_cast_fp16)[name = string("op_913_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_915_cast_fp16 = mul(x = var_913_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_915_cast_fp16")];
            int32 var_917 = const()[name = string("op_917"), val = int32(-2)];
            bool var_918_interleave_0 = const()[name = string("op_918_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_918_cast_fp16 = concat(axis = var_917, interleave = var_918_interleave_0, values = (var_915_cast_fp16, var_913_cast_fp16_0))[name = string("op_918_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_919_cast_fp16 = mul(x = var_918_cast_fp16, y = var_459_cast_fp16)[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_9_cast_fp16 = add(x = var_912_cast_fp16, y = var_919_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_925_cast_fp16 = mul(x = var_901_cast_fp16, y = var_452_cast_fp16)[name = string("op_925_cast_fp16")];
            tensor<int32, [2]> var_926_split_sizes_0 = const()[name = string("op_926_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_926_axis_0 = const()[name = string("op_926_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_926_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_926_cast_fp16_1 = split(axis = var_926_axis_0, split_sizes = var_926_split_sizes_0, x = var_901_cast_fp16)[name = string("op_926_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_928_cast_fp16 = mul(x = var_926_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_928_cast_fp16")];
            int32 var_930 = const()[name = string("op_930"), val = int32(-2)];
            bool var_931_interleave_0 = const()[name = string("op_931_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_931_cast_fp16 = concat(axis = var_930, interleave = var_931_interleave_0, values = (var_928_cast_fp16, var_926_cast_fp16_0))[name = string("op_931_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_932_cast_fp16 = mul(x = var_931_cast_fp16, y = var_459_cast_fp16)[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_15_cast_fp16 = add(x = var_925_cast_fp16, y = var_932_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_263")];
            tensor<fp16, [14, 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_140)[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_142_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_142 = read_state(input = key_cache)[name = string("coreml_update_state_142")];
            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_908_cast_fp16)[name = string("transpose_262")];
            tensor<fp16, [14, 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_141)[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_143_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_143 = read_state(input = value_cache)[name = string("coreml_update_state_143")];
            tensor<int32, [4]> var_1002_begin_0 = const()[name = string("op_1002_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1002_end_0 = const()[name = string("op_1002_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1002_end_mask_0 = const()[name = string("op_1002_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1002_cast_fp16 = slice_by_index(begin = var_1002_begin_0, end = var_1002_end_0, end_mask = var_1002_end_mask_0, x = coreml_update_state_142)[name = string("op_1002_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1005_axis_0 = const()[name = string("op_1005_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_1 = split(axis = var_1005_axis_0, split_sizes = tile_2, x = var_1002_cast_fp16)[name = string("op_1005_cast_fp16")];
            tensor<int32, [4]> var_1012_begin_0 = const()[name = string("op_1012_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1012_end_0 = const()[name = string("op_1012_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1012_end_mask_0 = const()[name = string("op_1012_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1012_cast_fp16 = slice_by_index(begin = var_1012_begin_0, end = var_1012_end_0, end_mask = var_1012_end_mask_0, x = coreml_update_state_143)[name = string("op_1012_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1015_axis_0 = const()[name = string("op_1015_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_1 = split(axis = var_1015_axis_0, split_sizes = tile_3, x = var_1012_cast_fp16)[name = string("op_1015_cast_fp16")];
            tensor<int32, [2]> var_1018_split_sizes_0 = const()[name = string("op_1018_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1018_axis_0 = const()[name = string("op_1018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1018_0, tensor<fp16, [1, 8, 128, 64]> var_1018_1 = split(axis = var_1018_axis_0, split_sizes = var_1018_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1018")];
            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_1005_cast_fp16_0, y = var_1018_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1021_to_fp16 = const()[name = string("op_1021_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_1021_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_1025 = const()[name = string("op_1025"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_23_cast_fp16 = softmax(axis = var_1025, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1031_transpose_x_1 = const()[name = string("op_1031_transpose_x_1"), val = bool(true)];
            bool var_1031_transpose_y_1 = const()[name = string("op_1031_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1031_cast_fp16 = matmul(transpose_x = var_1031_transpose_x_1, transpose_y = var_1031_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1015_cast_fp16_0)[name = string("op_1031_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_1005_cast_fp16_1, y = var_1018_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1033_to_fp16 = const()[name = string("op_1033_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_1033_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_1037 = const()[name = string("op_1037"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_31_cast_fp16 = softmax(axis = var_1037, 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_1015_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1045 = const()[name = string("op_1045"), 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_1045, interleave = attn_output_11_interleave_0, values = (var_1031_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1049_perm_0 = const()[name = string("op_1049_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_1049_cast_fp16 = transpose(perm = var_1049_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_261")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1049_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553154624)))];
            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_to_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_1082_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1082_cast_fp16")];
            int32 var_1080 = const()[name = string("op_1080"), 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_1080, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1082_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(561543296)))];
            fp16 var_1092_to_fp16 = const()[name = string("op_1092_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_1092_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1103_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1103_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561551552)))];
            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_to_fp16, x = var_1103_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_1120_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [2]> var_1126_strides_0 = const()[name = string("op_1126_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1126_pad_type_0 = const()[name = string("op_1126_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1126_pad_0 = const()[name = string("op_1126_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1126_dilations_0 = const()[name = string("op_1126_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1126_groups_0 = const()[name = string("op_1126_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_1126_cast_fp16 = conv(dilations = var_1126_dilations_0, groups = var_1126_groups_0, pad = var_1126_pad_0, pad_type = var_1126_pad_type_0, strides = var_1126_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1103_cast_fp16_0)[name = string("op_1126_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_19_cast_fp16 = mul(x = var_1120_cast_fp16, y = var_1126_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586717440)))];
            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_to_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_1144_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1144_cast_fp16")];
            int32 var_1142 = const()[name = string("op_1142"), 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_1142, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1144_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(611883328)))];
            fp16 var_1154_to_fp16 = const()[name = string("op_1154_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_1154_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1165_split_sizes_0 = const()[name = string("op_1165_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1165_axis_0 = const()[name = string("op_1165_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1165_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1165_cast_fp16_1 = split(axis = var_1165_axis_0, split_sizes = var_1165_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1165_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611891584)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620280256)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            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_cast_fp16, x = var_1165_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_1222_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1222_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_1229_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1229_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1233_cast_fp16 = mul(x = x_21_cast_fp16, y = var_452_cast_fp16)[name = string("op_1233_cast_fp16")];
            tensor<int32, [2]> var_1234_split_sizes_0 = const()[name = string("op_1234_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1234_axis_0 = const()[name = string("op_1234_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_1234_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1234_cast_fp16_1 = split(axis = var_1234_axis_0, split_sizes = var_1234_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1234_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_1236_cast_fp16 = mul(x = var_1234_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1236_cast_fp16")];
            int32 var_1238 = const()[name = string("op_1238"), val = int32(-2)];
            bool var_1239_interleave_0 = const()[name = string("op_1239_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1239_cast_fp16 = concat(axis = var_1238, interleave = var_1239_interleave_0, values = (var_1236_cast_fp16, var_1234_cast_fp16_0))[name = string("op_1239_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1240_cast_fp16 = mul(x = var_1239_cast_fp16, y = var_459_cast_fp16)[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_15_cast_fp16 = add(x = var_1233_cast_fp16, y = var_1240_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1246_cast_fp16 = mul(x = var_1222_cast_fp16, y = var_452_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<int32, [2]> var_1247_split_sizes_0 = const()[name = string("op_1247_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1247_axis_0 = const()[name = string("op_1247_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_1247_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_1247_cast_fp16_1 = split(axis = var_1247_axis_0, split_sizes = var_1247_split_sizes_0, x = var_1222_cast_fp16)[name = string("op_1247_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_1249_cast_fp16 = mul(x = var_1247_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1249_cast_fp16")];
            int32 var_1251 = const()[name = string("op_1251"), val = int32(-2)];
            bool var_1252_interleave_0 = const()[name = string("op_1252_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1252_cast_fp16 = concat(axis = var_1251, interleave = var_1252_interleave_0, values = (var_1249_cast_fp16, var_1247_cast_fp16_0))[name = string("op_1252_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1253_cast_fp16 = mul(x = var_1252_cast_fp16, y = var_459_cast_fp16)[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_25_cast_fp16 = add(x = var_1246_cast_fp16, y = var_1253_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_260")];
            tensor<fp16, [14, 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_142)[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_144_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_144 = read_state(input = key_cache)[name = string("coreml_update_state_144")];
            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_1229_cast_fp16)[name = string("transpose_259")];
            tensor<fp16, [14, 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_143)[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_145_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_145 = read_state(input = value_cache)[name = string("coreml_update_state_145")];
            tensor<int32, [4]> var_1323_begin_0 = const()[name = string("op_1323_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1323_end_0 = const()[name = string("op_1323_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1323_end_mask_0 = const()[name = string("op_1323_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1323_cast_fp16 = slice_by_index(begin = var_1323_begin_0, end = var_1323_end_0, end_mask = var_1323_end_mask_0, x = coreml_update_state_144)[name = string("op_1323_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1326_axis_0 = const()[name = string("op_1326_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_1 = split(axis = var_1326_axis_0, split_sizes = tile_4, x = var_1323_cast_fp16)[name = string("op_1326_cast_fp16")];
            tensor<int32, [4]> var_1333_begin_0 = const()[name = string("op_1333_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1333_end_0 = const()[name = string("op_1333_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1333_end_mask_0 = const()[name = string("op_1333_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1333_cast_fp16 = slice_by_index(begin = var_1333_begin_0, end = var_1333_end_0, end_mask = var_1333_end_mask_0, x = coreml_update_state_145)[name = string("op_1333_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1336_axis_0 = const()[name = string("op_1336_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_1 = split(axis = var_1336_axis_0, split_sizes = tile_5, x = var_1333_cast_fp16)[name = string("op_1336_cast_fp16")];
            tensor<int32, [2]> var_1339_split_sizes_0 = const()[name = string("op_1339_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1339_axis_0 = const()[name = string("op_1339_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1339_0, tensor<fp16, [1, 8, 128, 64]> var_1339_1 = split(axis = var_1339_axis_0, split_sizes = var_1339_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1339")];
            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_1326_cast_fp16_0, y = var_1339_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1342_to_fp16 = const()[name = string("op_1342_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_1342_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_1346 = const()[name = string("op_1346"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_39_cast_fp16 = softmax(axis = var_1346, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1352_transpose_x_1 = const()[name = string("op_1352_transpose_x_1"), val = bool(true)];
            bool var_1352_transpose_y_1 = const()[name = string("op_1352_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1352_cast_fp16 = matmul(transpose_x = var_1352_transpose_x_1, transpose_y = var_1352_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1336_cast_fp16_0)[name = string("op_1352_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_1326_cast_fp16_1, y = var_1339_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1354_to_fp16 = const()[name = string("op_1354_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_1354_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_1358 = const()[name = string("op_1358"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_47_cast_fp16 = softmax(axis = var_1358, 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_1336_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1366 = const()[name = string("op_1366"), 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_1366, interleave = attn_output_19_interleave_0, values = (var_1352_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1370_perm_0 = const()[name = string("op_1370_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_1370_cast_fp16 = transpose(perm = var_1370_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_258")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1370_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621328896)))];
            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_to_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_1403_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1403_cast_fp16")];
            int32 var_1401 = const()[name = string("op_1401"), 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_1401, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1403_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(629717568)))];
            fp16 var_1413_to_fp16 = const()[name = string("op_1413_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_1413_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1424_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1424_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629725824)))];
            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_to_fp16, x = var_1424_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_1441_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1441_cast_fp16")];
            tensor<int32, [2]> var_1447_strides_0 = const()[name = string("op_1447_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1447_pad_type_0 = const()[name = string("op_1447_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1447_pad_0 = const()[name = string("op_1447_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1447_dilations_0 = const()[name = string("op_1447_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1447_groups_0 = const()[name = string("op_1447_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_1447_cast_fp16 = conv(dilations = var_1447_dilations_0, groups = var_1447_groups_0, pad = var_1447_pad_0, pad_type = var_1447_pad_type_0, strides = var_1447_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1424_cast_fp16_0)[name = string("op_1447_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_29_cast_fp16 = mul(x = var_1441_cast_fp16, y = var_1447_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654891712)))];
            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_to_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_1465_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1465_cast_fp16")];
            int32 var_1463 = const()[name = string("op_1463"), 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_1463, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1465_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(680057600)))];
            fp16 var_1475_to_fp16 = const()[name = string("op_1475_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_1475_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1486_split_sizes_0 = const()[name = string("op_1486_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1486_axis_0 = const()[name = string("op_1486_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1486_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1486_cast_fp16_1 = split(axis = var_1486_axis_0, split_sizes = var_1486_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1486_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680065856)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688454528)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            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_cast_fp16, x = var_1486_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_1543_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1543_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_1550_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1554_cast_fp16 = mul(x = x_31_cast_fp16, y = var_452_cast_fp16)[name = string("op_1554_cast_fp16")];
            tensor<int32, [2]> var_1555_split_sizes_0 = const()[name = string("op_1555_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1555_axis_0 = const()[name = string("op_1555_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_1555_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1555_cast_fp16_1 = split(axis = var_1555_axis_0, split_sizes = var_1555_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1555_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_1557_cast_fp16 = mul(x = var_1555_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1557_cast_fp16")];
            int32 var_1559 = const()[name = string("op_1559"), val = int32(-2)];
            bool var_1560_interleave_0 = const()[name = string("op_1560_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1560_cast_fp16 = concat(axis = var_1559, interleave = var_1560_interleave_0, values = (var_1557_cast_fp16, var_1555_cast_fp16_0))[name = string("op_1560_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1561_cast_fp16 = mul(x = var_1560_cast_fp16, y = var_459_cast_fp16)[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_21_cast_fp16 = add(x = var_1554_cast_fp16, y = var_1561_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1567_cast_fp16 = mul(x = var_1543_cast_fp16, y = var_452_cast_fp16)[name = string("op_1567_cast_fp16")];
            tensor<int32, [2]> var_1568_split_sizes_0 = const()[name = string("op_1568_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1568_axis_0 = const()[name = string("op_1568_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_1568_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_1568_cast_fp16_1 = split(axis = var_1568_axis_0, split_sizes = var_1568_split_sizes_0, x = var_1543_cast_fp16)[name = string("op_1568_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_1570_cast_fp16 = mul(x = var_1568_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1570_cast_fp16")];
            int32 var_1572 = const()[name = string("op_1572"), val = int32(-2)];
            bool var_1573_interleave_0 = const()[name = string("op_1573_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1573_cast_fp16 = concat(axis = var_1572, interleave = var_1573_interleave_0, values = (var_1570_cast_fp16, var_1568_cast_fp16_0))[name = string("op_1573_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1574_cast_fp16 = mul(x = var_1573_cast_fp16, y = var_459_cast_fp16)[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_35_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1574_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_257")];
            tensor<fp16, [14, 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_144)[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_146_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_146 = read_state(input = key_cache)[name = string("coreml_update_state_146")];
            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_1550_cast_fp16)[name = string("transpose_256")];
            tensor<fp16, [14, 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_145)[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_147_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_147 = read_state(input = value_cache)[name = string("coreml_update_state_147")];
            tensor<int32, [4]> var_1644_begin_0 = const()[name = string("op_1644_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1644_end_0 = const()[name = string("op_1644_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1644_end_mask_0 = const()[name = string("op_1644_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1644_cast_fp16 = slice_by_index(begin = var_1644_begin_0, end = var_1644_end_0, end_mask = var_1644_end_mask_0, x = coreml_update_state_146)[name = string("op_1644_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1647_axis_0 = const()[name = string("op_1647_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_1 = split(axis = var_1647_axis_0, split_sizes = tile_6, x = var_1644_cast_fp16)[name = string("op_1647_cast_fp16")];
            tensor<int32, [4]> var_1654_begin_0 = const()[name = string("op_1654_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1654_end_0 = const()[name = string("op_1654_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1654_end_mask_0 = const()[name = string("op_1654_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1654_cast_fp16 = slice_by_index(begin = var_1654_begin_0, end = var_1654_end_0, end_mask = var_1654_end_mask_0, x = coreml_update_state_147)[name = string("op_1654_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1657_axis_0 = const()[name = string("op_1657_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_1 = split(axis = var_1657_axis_0, split_sizes = tile_7, x = var_1654_cast_fp16)[name = string("op_1657_cast_fp16")];
            tensor<int32, [2]> var_1660_split_sizes_0 = const()[name = string("op_1660_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1660_axis_0 = const()[name = string("op_1660_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1660_0, tensor<fp16, [1, 8, 128, 64]> var_1660_1 = split(axis = var_1660_axis_0, split_sizes = var_1660_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1660")];
            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_1647_cast_fp16_0, y = var_1660_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1663_to_fp16 = const()[name = string("op_1663_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_1663_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_1667 = const()[name = string("op_1667"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_55_cast_fp16 = softmax(axis = var_1667, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1673_transpose_x_1 = const()[name = string("op_1673_transpose_x_1"), val = bool(true)];
            bool var_1673_transpose_y_1 = const()[name = string("op_1673_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1673_cast_fp16 = matmul(transpose_x = var_1673_transpose_x_1, transpose_y = var_1673_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1657_cast_fp16_0)[name = string("op_1673_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_1647_cast_fp16_1, y = var_1660_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1675_to_fp16 = const()[name = string("op_1675_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_1675_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_1679 = const()[name = string("op_1679"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_63_cast_fp16 = softmax(axis = var_1679, 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_1657_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1687 = const()[name = string("op_1687"), 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_1687, interleave = attn_output_27_interleave_0, values = (var_1673_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1691_perm_0 = const()[name = string("op_1691_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_1691_cast_fp16 = transpose(perm = var_1691_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_255")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1691_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            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_cast_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_1724_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1724_cast_fp16")];
            int32 var_1722 = const()[name = string("op_1722"), 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_1722, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1724_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(689503168)))];
            fp16 var_1734_to_fp16 = const()[name = string("op_1734_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_1734_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1745_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1745_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689511424)))];
            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_to_fp16, x = var_1745_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_1762_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1762_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(714677312)))];
            tensor<int32, [2]> var_1768_strides_0 = const()[name = string("op_1768_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1768_pad_type_0 = const()[name = string("op_1768_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1768_pad_0 = const()[name = string("op_1768_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1768_dilations_0 = const()[name = string("op_1768_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1768_groups_0 = const()[name = string("op_1768_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_1768_cast_fp16 = conv(dilations = var_1768_dilations_0, groups = var_1768_groups_0, pad = var_1768_pad_0, pad_type = var_1768_pad_type_0, strides = var_1768_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1745_cast_fp16_0)[name = string("op_1768_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_39_cast_fp16 = mul(x = var_1762_cast_fp16, y = var_1768_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_1786_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1786_cast_fp16")];
            int32 var_1784 = const()[name = string("op_1784"), 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_1784, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1786_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(739843200)))];
            fp16 var_1796_to_fp16 = const()[name = string("op_1796_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_1796_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1807_split_sizes_0 = const()[name = string("op_1807_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1807_axis_0 = const()[name = string("op_1807_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_1807_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_1807_cast_fp16_1 = split(axis = var_1807_axis_0, split_sizes = var_1807_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739851456)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(748240128)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            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_cast_fp16, x = var_1807_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_1864_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1864_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_1871_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1871_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1875_cast_fp16 = mul(x = x_41_cast_fp16, y = var_452_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [2]> var_1876_split_sizes_0 = const()[name = string("op_1876_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1876_axis_0 = const()[name = string("op_1876_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_1876_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1876_cast_fp16_1 = split(axis = var_1876_axis_0, split_sizes = var_1876_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1876_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_1878_cast_fp16 = mul(x = var_1876_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1878_cast_fp16")];
            int32 var_1880 = const()[name = string("op_1880"), val = int32(-2)];
            bool var_1881_interleave_0 = const()[name = string("op_1881_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_1881_cast_fp16 = concat(axis = var_1880, interleave = var_1881_interleave_0, values = (var_1878_cast_fp16, var_1876_cast_fp16_0))[name = string("op_1881_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1882_cast_fp16 = mul(x = var_1881_cast_fp16, y = var_459_cast_fp16)[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_27_cast_fp16 = add(x = var_1875_cast_fp16, y = var_1882_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1888_cast_fp16 = mul(x = var_1864_cast_fp16, y = var_452_cast_fp16)[name = string("op_1888_cast_fp16")];
            tensor<int32, [2]> var_1889_split_sizes_0 = const()[name = string("op_1889_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1889_axis_0 = const()[name = string("op_1889_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_1889_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_1889_cast_fp16_1 = split(axis = var_1889_axis_0, split_sizes = var_1889_split_sizes_0, x = var_1864_cast_fp16)[name = string("op_1889_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_1891_cast_fp16 = mul(x = var_1889_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1891_cast_fp16")];
            int32 var_1893 = const()[name = string("op_1893"), val = int32(-2)];
            bool var_1894_interleave_0 = const()[name = string("op_1894_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1894_cast_fp16 = concat(axis = var_1893, interleave = var_1894_interleave_0, values = (var_1891_cast_fp16, var_1889_cast_fp16_0))[name = string("op_1894_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1895_cast_fp16 = mul(x = var_1894_cast_fp16, y = var_459_cast_fp16)[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_45_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1895_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_254")];
            tensor<fp16, [14, 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_146)[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_148_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_148 = read_state(input = key_cache)[name = string("coreml_update_state_148")];
            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_1871_cast_fp16)[name = string("transpose_253")];
            tensor<fp16, [14, 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_147)[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_149_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_149 = read_state(input = value_cache)[name = string("coreml_update_state_149")];
            tensor<int32, [4]> var_1965_begin_0 = const()[name = string("op_1965_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1965_end_0 = const()[name = string("op_1965_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1965_end_mask_0 = const()[name = string("op_1965_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1965_cast_fp16 = slice_by_index(begin = var_1965_begin_0, end = var_1965_end_0, end_mask = var_1965_end_mask_0, x = coreml_update_state_148)[name = string("op_1965_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1968_axis_0 = const()[name = string("op_1968_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_1 = split(axis = var_1968_axis_0, split_sizes = tile_8, x = var_1965_cast_fp16)[name = string("op_1968_cast_fp16")];
            tensor<int32, [4]> var_1975_begin_0 = const()[name = string("op_1975_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1975_end_0 = const()[name = string("op_1975_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1975_end_mask_0 = const()[name = string("op_1975_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1975_cast_fp16 = slice_by_index(begin = var_1975_begin_0, end = var_1975_end_0, end_mask = var_1975_end_mask_0, x = coreml_update_state_149)[name = string("op_1975_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1978_axis_0 = const()[name = string("op_1978_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_1 = split(axis = var_1978_axis_0, split_sizes = tile_9, x = var_1975_cast_fp16)[name = string("op_1978_cast_fp16")];
            tensor<int32, [2]> var_1981_split_sizes_0 = const()[name = string("op_1981_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1981_axis_0 = const()[name = string("op_1981_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_1981_0, tensor<fp16, [1, 8, 128, 64]> var_1981_1 = split(axis = var_1981_axis_0, split_sizes = var_1981_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1981")];
            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_1968_cast_fp16_0, y = var_1981_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1984_to_fp16 = const()[name = string("op_1984_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_1984_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_1988 = const()[name = string("op_1988"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_71_cast_fp16 = softmax(axis = var_1988, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1994_transpose_x_1 = const()[name = string("op_1994_transpose_x_1"), val = bool(true)];
            bool var_1994_transpose_y_1 = const()[name = string("op_1994_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1994_cast_fp16 = matmul(transpose_x = var_1994_transpose_x_1, transpose_y = var_1994_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1978_cast_fp16_0)[name = string("op_1994_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_1968_cast_fp16_1, y = var_1981_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1996_to_fp16 = const()[name = string("op_1996_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_1996_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_2000 = const()[name = string("op_2000"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_79_cast_fp16 = softmax(axis = var_2000, 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_1978_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2008 = const()[name = string("op_2008"), 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_2008, interleave = attn_output_35_interleave_0, values = (var_1994_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2012_perm_0 = const()[name = string("op_2012_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_2012_cast_fp16 = transpose(perm = var_2012_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_252")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2012_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_2045_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2045_cast_fp16")];
            int32 var_2043 = const()[name = string("op_2043"), 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_2043, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2045_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(749288768)))];
            fp16 var_2055_to_fp16 = const()[name = string("op_2055_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_2055_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2066_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2066_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_2066_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_2083_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2083_cast_fp16")];
            tensor<int32, [2]> var_2089_strides_0 = const()[name = string("op_2089_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2089_pad_type_0 = const()[name = string("op_2089_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2089_pad_0 = const()[name = string("op_2089_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2089_dilations_0 = const()[name = string("op_2089_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2089_groups_0 = const()[name = string("op_2089_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_2089_cast_fp16 = conv(dilations = var_2089_dilations_0, groups = var_2089_groups_0, pad = var_2089_pad_0, pad_type = var_2089_pad_type_0, strides = var_2089_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2066_cast_fp16_0)[name = string("op_2089_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_49_cast_fp16 = mul(x = var_2083_cast_fp16, y = var_2089_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_2107_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2107_cast_fp16")];
            int32 var_2105 = const()[name = string("op_2105"), 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_2105, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2107_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(749297024)))];
            fp16 var_2117_to_fp16 = const()[name = string("op_2117_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_2117_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2128_split_sizes_0 = const()[name = string("op_2128_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2128_axis_0 = const()[name = string("op_2128_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2128_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2128_cast_fp16_1 = split(axis = var_2128_axis_0, split_sizes = var_2128_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749305280)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(757693952)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            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_cast_fp16, x = var_2128_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_2185_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2185_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_2192_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2196_cast_fp16 = mul(x = x_51_cast_fp16, y = var_452_cast_fp16)[name = string("op_2196_cast_fp16")];
            tensor<int32, [2]> var_2197_split_sizes_0 = const()[name = string("op_2197_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2197_axis_0 = const()[name = string("op_2197_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_2197_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_2197_cast_fp16_1 = split(axis = var_2197_axis_0, split_sizes = var_2197_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2197_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_2199_cast_fp16 = mul(x = var_2197_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2199_cast_fp16")];
            int32 var_2201 = const()[name = string("op_2201"), val = int32(-2)];
            bool var_2202_interleave_0 = const()[name = string("op_2202_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2202_cast_fp16 = concat(axis = var_2201, interleave = var_2202_interleave_0, values = (var_2199_cast_fp16, var_2197_cast_fp16_0))[name = string("op_2202_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2203_cast_fp16 = mul(x = var_2202_cast_fp16, y = var_459_cast_fp16)[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_33_cast_fp16 = add(x = var_2196_cast_fp16, y = var_2203_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2209_cast_fp16 = mul(x = var_2185_cast_fp16, y = var_452_cast_fp16)[name = string("op_2209_cast_fp16")];
            tensor<int32, [2]> var_2210_split_sizes_0 = const()[name = string("op_2210_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2210_axis_0 = const()[name = string("op_2210_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_2210_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_2210_cast_fp16_1 = split(axis = var_2210_axis_0, split_sizes = var_2210_split_sizes_0, x = var_2185_cast_fp16)[name = string("op_2210_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_2212_cast_fp16 = mul(x = var_2210_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2212_cast_fp16")];
            int32 var_2214 = const()[name = string("op_2214"), val = int32(-2)];
            bool var_2215_interleave_0 = const()[name = string("op_2215_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2215_cast_fp16 = concat(axis = var_2214, interleave = var_2215_interleave_0, values = (var_2212_cast_fp16, var_2210_cast_fp16_0))[name = string("op_2215_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2216_cast_fp16 = mul(x = var_2215_cast_fp16, y = var_459_cast_fp16)[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_55_cast_fp16 = add(x = var_2209_cast_fp16, y = var_2216_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_251")];
            tensor<fp16, [14, 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_148)[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_150_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_150 = read_state(input = key_cache)[name = string("coreml_update_state_150")];
            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_2192_cast_fp16)[name = string("transpose_250")];
            tensor<fp16, [14, 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_149)[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_151_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_151 = read_state(input = value_cache)[name = string("coreml_update_state_151")];
            tensor<int32, [4]> var_2286_begin_0 = const()[name = string("op_2286_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2286_end_0 = const()[name = string("op_2286_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2286_end_mask_0 = const()[name = string("op_2286_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2286_cast_fp16 = slice_by_index(begin = var_2286_begin_0, end = var_2286_end_0, end_mask = var_2286_end_mask_0, x = coreml_update_state_150)[name = string("op_2286_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2289_axis_0 = const()[name = string("op_2289_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_1 = split(axis = var_2289_axis_0, split_sizes = tile_10, x = var_2286_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> var_2296_begin_0 = const()[name = string("op_2296_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2296_end_0 = const()[name = string("op_2296_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2296_end_mask_0 = const()[name = string("op_2296_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2296_cast_fp16 = slice_by_index(begin = var_2296_begin_0, end = var_2296_end_0, end_mask = var_2296_end_mask_0, x = coreml_update_state_151)[name = string("op_2296_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2299_axis_0 = const()[name = string("op_2299_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_1 = split(axis = var_2299_axis_0, split_sizes = tile_11, x = var_2296_cast_fp16)[name = string("op_2299_cast_fp16")];
            tensor<int32, [2]> var_2302_split_sizes_0 = const()[name = string("op_2302_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2302_axis_0 = const()[name = string("op_2302_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_2302_0, tensor<fp16, [1, 8, 128, 64]> var_2302_1 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2302")];
            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_2289_cast_fp16_0, y = var_2302_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2305_to_fp16 = const()[name = string("op_2305_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_2305_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_2309 = const()[name = string("op_2309"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_87_cast_fp16 = softmax(axis = var_2309, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2315_transpose_x_1 = const()[name = string("op_2315_transpose_x_1"), val = bool(true)];
            bool var_2315_transpose_y_1 = const()[name = string("op_2315_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_2315_cast_fp16 = matmul(transpose_x = var_2315_transpose_x_1, transpose_y = var_2315_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2299_cast_fp16_0)[name = string("op_2315_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_2289_cast_fp16_1, y = var_2302_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2317_to_fp16 = const()[name = string("op_2317_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_2317_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_2321 = const()[name = string("op_2321"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_95_cast_fp16 = softmax(axis = var_2321, 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_2299_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2329 = const()[name = string("op_2329"), 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_2329, interleave = attn_output_43_interleave_0, values = (var_2315_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2333_perm_0 = const()[name = string("op_2333_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_2333_cast_fp16 = transpose(perm = var_2333_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_249")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2333_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_2366_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2366_cast_fp16")];
            int32 var_2364 = const()[name = string("op_2364"), 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_2364, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2366_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(758742592)))];
            fp16 var_2376_to_fp16 = const()[name = string("op_2376_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_2376_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2387_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2387_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(758750848)))];
            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_to_fp16, x = var_2387_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_2404_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [2]> var_2410_strides_0 = const()[name = string("op_2410_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2410_pad_type_0 = const()[name = string("op_2410_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2410_pad_0 = const()[name = string("op_2410_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2410_dilations_0 = const()[name = string("op_2410_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2410_groups_0 = const()[name = string("op_2410_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_2410_cast_fp16 = conv(dilations = var_2410_dilations_0, groups = var_2410_groups_0, pad = var_2410_pad_0, pad_type = var_2410_pad_type_0, strides = var_2410_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2387_cast_fp16_0)[name = string("op_2410_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_59_cast_fp16 = mul(x = var_2404_cast_fp16, y = var_2410_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_2428_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2428_cast_fp16")];
            int32 var_2426 = const()[name = string("op_2426"), 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_2426, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2428_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(783916736)))];
            fp16 var_2438_to_fp16 = const()[name = string("op_2438_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_2438_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2449_split_sizes_0 = const()[name = string("op_2449_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2449_axis_0 = const()[name = string("op_2449_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2449_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2449_cast_fp16_1 = split(axis = var_2449_axis_0, split_sizes = var_2449_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2449_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(783924992)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792313664)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            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_cast_fp16, x = var_2449_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_2506_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2506_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_2513_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2517_cast_fp16 = mul(x = x_61_cast_fp16, y = var_452_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<int32, [2]> var_2518_split_sizes_0 = const()[name = string("op_2518_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2518_axis_0 = const()[name = string("op_2518_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_2518_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_2518_cast_fp16_1 = split(axis = var_2518_axis_0, split_sizes = var_2518_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2518_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_2520_cast_fp16 = mul(x = var_2518_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2520_cast_fp16")];
            int32 var_2522 = const()[name = string("op_2522"), val = int32(-2)];
            bool var_2523_interleave_0 = const()[name = string("op_2523_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2523_cast_fp16 = concat(axis = var_2522, interleave = var_2523_interleave_0, values = (var_2520_cast_fp16, var_2518_cast_fp16_0))[name = string("op_2523_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2524_cast_fp16 = mul(x = var_2523_cast_fp16, y = var_459_cast_fp16)[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_39_cast_fp16 = add(x = var_2517_cast_fp16, y = var_2524_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2530_cast_fp16 = mul(x = var_2506_cast_fp16, y = var_452_cast_fp16)[name = string("op_2530_cast_fp16")];
            tensor<int32, [2]> var_2531_split_sizes_0 = const()[name = string("op_2531_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2531_axis_0 = const()[name = string("op_2531_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_2531_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_2531_cast_fp16_1 = split(axis = var_2531_axis_0, split_sizes = var_2531_split_sizes_0, x = var_2506_cast_fp16)[name = string("op_2531_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_2533_cast_fp16 = mul(x = var_2531_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2533_cast_fp16")];
            int32 var_2535 = const()[name = string("op_2535"), val = int32(-2)];
            bool var_2536_interleave_0 = const()[name = string("op_2536_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2536_cast_fp16 = concat(axis = var_2535, interleave = var_2536_interleave_0, values = (var_2533_cast_fp16, var_2531_cast_fp16_0))[name = string("op_2536_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2537_cast_fp16 = mul(x = var_2536_cast_fp16, y = var_459_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_65_cast_fp16 = add(x = var_2530_cast_fp16, y = var_2537_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_248")];
            tensor<fp16, [14, 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_150)[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_152_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_152 = read_state(input = key_cache)[name = string("coreml_update_state_152")];
            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_2513_cast_fp16)[name = string("transpose_247")];
            tensor<fp16, [14, 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_151)[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_153_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_153 = read_state(input = value_cache)[name = string("coreml_update_state_153")];
            tensor<int32, [4]> var_2607_begin_0 = const()[name = string("op_2607_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2607_end_0 = const()[name = string("op_2607_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2607_end_mask_0 = const()[name = string("op_2607_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2607_cast_fp16 = slice_by_index(begin = var_2607_begin_0, end = var_2607_end_0, end_mask = var_2607_end_mask_0, x = coreml_update_state_152)[name = string("op_2607_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2610_axis_0 = const()[name = string("op_2610_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_1 = split(axis = var_2610_axis_0, split_sizes = tile_12, x = var_2607_cast_fp16)[name = string("op_2610_cast_fp16")];
            tensor<int32, [4]> var_2617_begin_0 = const()[name = string("op_2617_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2617_end_0 = const()[name = string("op_2617_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2617_end_mask_0 = const()[name = string("op_2617_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2617_cast_fp16 = slice_by_index(begin = var_2617_begin_0, end = var_2617_end_0, end_mask = var_2617_end_mask_0, x = coreml_update_state_153)[name = string("op_2617_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2620_axis_0 = const()[name = string("op_2620_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_1 = split(axis = var_2620_axis_0, split_sizes = tile_13, x = var_2617_cast_fp16)[name = string("op_2620_cast_fp16")];
            tensor<int32, [2]> var_2623_split_sizes_0 = const()[name = string("op_2623_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2623_axis_0 = const()[name = string("op_2623_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_2623_0, tensor<fp16, [1, 8, 128, 64]> var_2623_1 = split(axis = var_2623_axis_0, split_sizes = var_2623_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2623")];
            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_2610_cast_fp16_0, y = var_2623_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2626_to_fp16 = const()[name = string("op_2626_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_2626_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_2630 = const()[name = string("op_2630"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_103_cast_fp16 = softmax(axis = var_2630, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2636_transpose_x_1 = const()[name = string("op_2636_transpose_x_1"), val = bool(true)];
            bool var_2636_transpose_y_1 = const()[name = string("op_2636_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_2636_cast_fp16 = matmul(transpose_x = var_2636_transpose_x_1, transpose_y = var_2636_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2620_cast_fp16_0)[name = string("op_2636_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_2610_cast_fp16_1, y = var_2623_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2638_to_fp16 = const()[name = string("op_2638_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_2638_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_2642 = const()[name = string("op_2642"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_111_cast_fp16 = softmax(axis = var_2642, 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_2620_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2650 = const()[name = string("op_2650"), 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_2650, interleave = attn_output_51_interleave_0, values = (var_2636_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2654_perm_0 = const()[name = string("op_2654_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_2654_cast_fp16 = transpose(perm = var_2654_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_246")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2654_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_2687_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2687_cast_fp16")];
            int32 var_2685 = const()[name = string("op_2685"), 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_2685, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2687_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(793362304)))];
            fp16 var_2697_to_fp16 = const()[name = string("op_2697_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_2697_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2708_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2708_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_2708_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_2725_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2725_cast_fp16")];
            tensor<int32, [2]> var_2731_strides_0 = const()[name = string("op_2731_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2731_pad_type_0 = const()[name = string("op_2731_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2731_pad_0 = const()[name = string("op_2731_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2731_dilations_0 = const()[name = string("op_2731_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2731_groups_0 = const()[name = string("op_2731_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_2731_cast_fp16 = conv(dilations = var_2731_dilations_0, groups = var_2731_groups_0, pad = var_2731_pad_0, pad_type = var_2731_pad_type_0, strides = var_2731_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2708_cast_fp16_0)[name = string("op_2731_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_69_cast_fp16 = mul(x = var_2725_cast_fp16, y = var_2731_cast_fp16)[name = string("x_69_cast_fp16")];
            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_cast_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_2749_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2749_cast_fp16")];
            int32 var_2747 = const()[name = string("op_2747"), 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_2747, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2749_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(793370560)))];
            fp16 var_2759_to_fp16 = const()[name = string("op_2759_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_2759_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2770_split_sizes_0 = const()[name = string("op_2770_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2770_axis_0 = const()[name = string("op_2770_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_2770_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_2770_cast_fp16_1 = split(axis = var_2770_axis_0, split_sizes = var_2770_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2770_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(793378816)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(801767488)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            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_cast_fp16, x = var_2770_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_2827_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2827_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_2834_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2834_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2838_cast_fp16 = mul(x = x_71_cast_fp16, y = var_452_cast_fp16)[name = string("op_2838_cast_fp16")];
            tensor<int32, [2]> var_2839_split_sizes_0 = const()[name = string("op_2839_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2839_axis_0 = const()[name = string("op_2839_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_2839_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_2839_cast_fp16_1 = split(axis = var_2839_axis_0, split_sizes = var_2839_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2839_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_2841_cast_fp16 = mul(x = var_2839_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2841_cast_fp16")];
            int32 var_2843 = const()[name = string("op_2843"), val = int32(-2)];
            bool var_2844_interleave_0 = const()[name = string("op_2844_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_2844_cast_fp16 = concat(axis = var_2843, interleave = var_2844_interleave_0, values = (var_2841_cast_fp16, var_2839_cast_fp16_0))[name = string("op_2844_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2845_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_459_cast_fp16)[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_45_cast_fp16 = add(x = var_2838_cast_fp16, y = var_2845_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2851_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_452_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [2]> var_2852_split_sizes_0 = const()[name = string("op_2852_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2852_axis_0 = const()[name = string("op_2852_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_2852_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_2852_cast_fp16_1 = split(axis = var_2852_axis_0, split_sizes = var_2852_split_sizes_0, x = var_2827_cast_fp16)[name = string("op_2852_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_2854_cast_fp16 = mul(x = var_2852_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2854_cast_fp16")];
            int32 var_2856 = const()[name = string("op_2856"), val = int32(-2)];
            bool var_2857_interleave_0 = const()[name = string("op_2857_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2857_cast_fp16 = concat(axis = var_2856, interleave = var_2857_interleave_0, values = (var_2854_cast_fp16, var_2852_cast_fp16_0))[name = string("op_2857_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2858_cast_fp16 = mul(x = var_2857_cast_fp16, y = var_459_cast_fp16)[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_75_cast_fp16 = add(x = var_2851_cast_fp16, y = var_2858_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_245")];
            tensor<fp16, [14, 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_152)[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_154_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_154 = read_state(input = key_cache)[name = string("coreml_update_state_154")];
            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_2834_cast_fp16)[name = string("transpose_244")];
            tensor<fp16, [14, 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_153)[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_155_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_155 = read_state(input = value_cache)[name = string("coreml_update_state_155")];
            tensor<int32, [4]> var_2928_begin_0 = const()[name = string("op_2928_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2928_end_0 = const()[name = string("op_2928_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2928_end_mask_0 = const()[name = string("op_2928_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2928_cast_fp16 = slice_by_index(begin = var_2928_begin_0, end = var_2928_end_0, end_mask = var_2928_end_mask_0, x = coreml_update_state_154)[name = string("op_2928_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2931_axis_0 = const()[name = string("op_2931_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_1 = split(axis = var_2931_axis_0, split_sizes = tile_14, x = var_2928_cast_fp16)[name = string("op_2931_cast_fp16")];
            tensor<int32, [4]> var_2938_begin_0 = const()[name = string("op_2938_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2938_end_0 = const()[name = string("op_2938_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2938_end_mask_0 = const()[name = string("op_2938_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2938_cast_fp16 = slice_by_index(begin = var_2938_begin_0, end = var_2938_end_0, end_mask = var_2938_end_mask_0, x = coreml_update_state_155)[name = string("op_2938_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2941_axis_0 = const()[name = string("op_2941_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_1 = split(axis = var_2941_axis_0, split_sizes = tile_15, x = var_2938_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<int32, [2]> var_2944_split_sizes_0 = const()[name = string("op_2944_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2944_axis_0 = const()[name = string("op_2944_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_2944_0, tensor<fp16, [1, 8, 128, 64]> var_2944_1 = split(axis = var_2944_axis_0, split_sizes = var_2944_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2944")];
            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_2931_cast_fp16_0, y = var_2944_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2947_to_fp16 = const()[name = string("op_2947_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_2947_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_2951 = const()[name = string("op_2951"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_119_cast_fp16 = softmax(axis = var_2951, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2957_transpose_x_1 = const()[name = string("op_2957_transpose_x_1"), val = bool(true)];
            bool var_2957_transpose_y_1 = const()[name = string("op_2957_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_2957_cast_fp16 = matmul(transpose_x = var_2957_transpose_x_1, transpose_y = var_2957_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2941_cast_fp16_0)[name = string("op_2957_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_2931_cast_fp16_1, y = var_2944_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2959_to_fp16 = const()[name = string("op_2959_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_2959_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_2963 = const()[name = string("op_2963"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_127_cast_fp16 = softmax(axis = var_2963, 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_2941_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2971 = const()[name = string("op_2971"), 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_2971, interleave = attn_output_59_interleave_0, values = (var_2957_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2975_perm_0 = const()[name = string("op_2975_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_2975_cast_fp16 = transpose(perm = var_2975_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_243")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2975_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_3008_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3008_cast_fp16")];
            int32 var_3006 = const()[name = string("op_3006"), 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_3006, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3008_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(802816128)))];
            fp16 var_3018_to_fp16 = const()[name = string("op_3018_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_3018_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3029_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3029_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_3029_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_3046_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3046_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(802824384)))];
            tensor<int32, [2]> var_3052_strides_0 = const()[name = string("op_3052_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3052_pad_type_0 = const()[name = string("op_3052_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3052_pad_0 = const()[name = string("op_3052_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3052_dilations_0 = const()[name = string("op_3052_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3052_groups_0 = const()[name = string("op_3052_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_3052_cast_fp16 = conv(dilations = var_3052_dilations_0, groups = var_3052_groups_0, pad = var_3052_pad_0, pad_type = var_3052_pad_type_0, strides = var_3052_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_3029_cast_fp16_0)[name = string("op_3052_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_79_cast_fp16 = mul(x = var_3046_cast_fp16, y = var_3052_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(827990272)))];
            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_3070_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3070_cast_fp16")];
            int32 var_3068 = const()[name = string("op_3068"), 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_3068, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3070_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(853156160)))];
            fp16 var_3080_to_fp16 = const()[name = string("op_3080_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_3080_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3091_split_sizes_0 = const()[name = string("op_3091_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3091_axis_0 = const()[name = string("op_3091_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3091_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3091_cast_fp16_1 = split(axis = var_3091_axis_0, split_sizes = var_3091_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3091_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(853164416)))];
            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_3091_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(861553088)))];
            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_3091_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            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_cast_fp16, x = var_3091_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_3148_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3148_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_3155_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3155_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3159_cast_fp16 = mul(x = x_81_cast_fp16, y = var_452_cast_fp16)[name = string("op_3159_cast_fp16")];
            tensor<int32, [2]> var_3160_split_sizes_0 = const()[name = string("op_3160_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3160_axis_0 = const()[name = string("op_3160_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_3160_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_3160_cast_fp16_1 = split(axis = var_3160_axis_0, split_sizes = var_3160_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3160_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_3162_cast_fp16 = mul(x = var_3160_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3162_cast_fp16")];
            int32 var_3164 = const()[name = string("op_3164"), val = int32(-2)];
            bool var_3165_interleave_0 = const()[name = string("op_3165_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_3165_cast_fp16 = concat(axis = var_3164, interleave = var_3165_interleave_0, values = (var_3162_cast_fp16, var_3160_cast_fp16_0))[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3166_cast_fp16 = mul(x = var_3165_cast_fp16, y = var_459_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_51_cast_fp16 = add(x = var_3159_cast_fp16, y = var_3166_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3172_cast_fp16 = mul(x = var_3148_cast_fp16, y = var_452_cast_fp16)[name = string("op_3172_cast_fp16")];
            tensor<int32, [2]> var_3173_split_sizes_0 = const()[name = string("op_3173_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3173_axis_0 = const()[name = string("op_3173_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_3173_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_3173_cast_fp16_1 = split(axis = var_3173_axis_0, split_sizes = var_3173_split_sizes_0, x = var_3148_cast_fp16)[name = string("op_3173_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_3175_cast_fp16 = mul(x = var_3173_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3175_cast_fp16")];
            int32 var_3177 = const()[name = string("op_3177"), val = int32(-2)];
            bool var_3178_interleave_0 = const()[name = string("op_3178_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3178_cast_fp16 = concat(axis = var_3177, interleave = var_3178_interleave_0, values = (var_3175_cast_fp16, var_3173_cast_fp16_0))[name = string("op_3178_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3179_cast_fp16 = mul(x = var_3178_cast_fp16, y = var_459_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_85_cast_fp16 = add(x = var_3172_cast_fp16, y = var_3179_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_242")];
            tensor<fp16, [14, 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_154)[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_156_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_156 = read_state(input = key_cache)[name = string("coreml_update_state_156")];
            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_3155_cast_fp16)[name = string("transpose_241")];
            tensor<fp16, [14, 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_155)[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_157_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_157 = read_state(input = value_cache)[name = string("coreml_update_state_157")];
            tensor<int32, [4]> var_3249_begin_0 = const()[name = string("op_3249_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3249_end_0 = const()[name = string("op_3249_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3249_end_mask_0 = const()[name = string("op_3249_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3249_cast_fp16 = slice_by_index(begin = var_3249_begin_0, end = var_3249_end_0, end_mask = var_3249_end_mask_0, x = coreml_update_state_156)[name = string("op_3249_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3252_axis_0 = const()[name = string("op_3252_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_1 = split(axis = var_3252_axis_0, split_sizes = tile_16, x = var_3249_cast_fp16)[name = string("op_3252_cast_fp16")];
            tensor<int32, [4]> var_3259_begin_0 = const()[name = string("op_3259_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3259_end_0 = const()[name = string("op_3259_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3259_end_mask_0 = const()[name = string("op_3259_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3259_cast_fp16 = slice_by_index(begin = var_3259_begin_0, end = var_3259_end_0, end_mask = var_3259_end_mask_0, x = coreml_update_state_157)[name = string("op_3259_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3262_axis_0 = const()[name = string("op_3262_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_1 = split(axis = var_3262_axis_0, split_sizes = tile_17, x = var_3259_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<int32, [2]> var_3265_split_sizes_0 = const()[name = string("op_3265_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3265_axis_0 = const()[name = string("op_3265_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_3265_0, tensor<fp16, [1, 8, 128, 64]> var_3265_1 = split(axis = var_3265_axis_0, split_sizes = var_3265_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3265")];
            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_3252_cast_fp16_0, y = var_3265_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3268_to_fp16 = const()[name = string("op_3268_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_3268_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_3272 = const()[name = string("op_3272"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_135_cast_fp16 = softmax(axis = var_3272, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3278_transpose_x_1 = const()[name = string("op_3278_transpose_x_1"), val = bool(true)];
            bool var_3278_transpose_y_1 = const()[name = string("op_3278_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_3278_cast_fp16 = matmul(transpose_x = var_3278_transpose_x_1, transpose_y = var_3278_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3262_cast_fp16_0)[name = string("op_3278_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_3252_cast_fp16_1, y = var_3265_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3280_to_fp16 = const()[name = string("op_3280_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_3280_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_3284 = const()[name = string("op_3284"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_143_cast_fp16 = softmax(axis = var_3284, x = attn_weights_141_cast_fp16)[name = string("attn_weights_143_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_143_cast_fp16, y = var_3262_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3292 = const()[name = string("op_3292"), 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_3292, interleave = attn_output_67_interleave_0, values = (var_3278_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3296_perm_0 = const()[name = string("op_3296_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_3296_cast_fp16 = transpose(perm = var_3296_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_240")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3296_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            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_cast_fp16, x = attn_output_71_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_3329_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3329_cast_fp16")];
            int32 var_3327 = const()[name = string("op_3327"), 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_3327, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3329_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(862601728)))];
            fp16 var_3339_to_fp16 = const()[name = string("op_3339_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_3339_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3350_split_sizes_0 = const()[name = string("op_3350_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3350_axis_0 = const()[name = string("op_3350_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3350_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3350_cast_fp16_1 = split(axis = var_3350_axis_0, split_sizes = var_3350_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3350_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_3367_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3367_cast_fp16")];
            tensor<int32, [2]> var_3373_strides_0 = const()[name = string("op_3373_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3373_pad_type_0 = const()[name = string("op_3373_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3373_pad_0 = const()[name = string("op_3373_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3373_dilations_0 = const()[name = string("op_3373_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3373_groups_0 = const()[name = string("op_3373_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_3373_cast_fp16 = conv(dilations = var_3373_dilations_0, groups = var_3373_groups_0, pad = var_3373_pad_0, pad_type = var_3373_pad_type_0, strides = var_3373_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("op_3373_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_89_cast_fp16 = mul(x = var_3367_cast_fp16, y = var_3373_cast_fp16)[name = string("x_89_cast_fp16")];
            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_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_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_3391_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3391_cast_fp16")];
            int32 var_3389 = const()[name = string("op_3389"), 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_3389, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3391_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_37_gamma_0_to_fp16 = const()[name = string("out_37_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862609984)))];
            fp16 var_3401_to_fp16 = const()[name = string("op_3401_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_3401_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3412_split_sizes_0 = const()[name = string("op_3412_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3412_axis_0 = const()[name = string("op_3412_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3412_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3412_cast_fp16_1 = split(axis = var_3412_axis_0, split_sizes = var_3412_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3412_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862618240)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(871006912)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3412_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_3469_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3469_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_3476_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3480_cast_fp16 = mul(x = x_91_cast_fp16, y = var_452_cast_fp16)[name = string("op_3480_cast_fp16")];
            tensor<int32, [2]> var_3481_split_sizes_0 = const()[name = string("op_3481_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3481_axis_0 = const()[name = string("op_3481_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_3481_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_3481_cast_fp16_1 = split(axis = var_3481_axis_0, split_sizes = var_3481_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3481_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_3483_cast_fp16 = mul(x = var_3481_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3483_cast_fp16")];
            int32 var_3485 = const()[name = string("op_3485"), val = int32(-2)];
            bool var_3486_interleave_0 = const()[name = string("op_3486_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_3486_cast_fp16 = concat(axis = var_3485, interleave = var_3486_interleave_0, values = (var_3483_cast_fp16, var_3481_cast_fp16_0))[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3487_cast_fp16 = mul(x = var_3486_cast_fp16, y = var_459_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_57_cast_fp16 = add(x = var_3480_cast_fp16, y = var_3487_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3493_cast_fp16 = mul(x = var_3469_cast_fp16, y = var_452_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [2]> var_3494_split_sizes_0 = const()[name = string("op_3494_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3494_axis_0 = const()[name = string("op_3494_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_3494_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_3494_cast_fp16_1 = split(axis = var_3494_axis_0, split_sizes = var_3494_split_sizes_0, x = var_3469_cast_fp16)[name = string("op_3494_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_3496_cast_fp16 = mul(x = var_3494_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3496_cast_fp16")];
            int32 var_3498 = const()[name = string("op_3498"), val = int32(-2)];
            bool var_3499_interleave_0 = const()[name = string("op_3499_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3499_cast_fp16 = concat(axis = var_3498, interleave = var_3499_interleave_0, values = (var_3496_cast_fp16, var_3494_cast_fp16_0))[name = string("op_3499_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3500_cast_fp16 = mul(x = var_3499_cast_fp16, y = var_459_cast_fp16)[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_95_cast_fp16 = add(x = var_3493_cast_fp16, y = var_3500_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_239")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_156)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_158_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_158 = read_state(input = key_cache)[name = string("coreml_update_state_158")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3476_cast_fp16)[name = string("transpose_238")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_157)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_159_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_159 = read_state(input = value_cache)[name = string("coreml_update_state_159")];
            tensor<int32, [4]> var_3570_begin_0 = const()[name = string("op_3570_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3570_end_0 = const()[name = string("op_3570_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3570_end_mask_0 = const()[name = string("op_3570_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = coreml_update_state_158)[name = string("op_3570_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3573_axis_0 = const()[name = string("op_3573_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_1 = split(axis = var_3573_axis_0, split_sizes = tile_18, x = var_3570_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [4]> var_3580_begin_0 = const()[name = string("op_3580_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3580_end_0 = const()[name = string("op_3580_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3580_end_mask_0 = const()[name = string("op_3580_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3580_cast_fp16 = slice_by_index(begin = var_3580_begin_0, end = var_3580_end_0, end_mask = var_3580_end_mask_0, x = coreml_update_state_159)[name = string("op_3580_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3583_axis_0 = const()[name = string("op_3583_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_1 = split(axis = var_3583_axis_0, split_sizes = tile_19, x = var_3580_cast_fp16)[name = string("op_3583_cast_fp16")];
            tensor<int32, [2]> var_3586_split_sizes_0 = const()[name = string("op_3586_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3586_axis_0 = const()[name = string("op_3586_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_3586_0, tensor<fp16, [1, 8, 128, 64]> var_3586_1 = split(axis = var_3586_axis_0, split_sizes = var_3586_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3586")];
            bool attn_weights_145_transpose_x_0 = const()[name = string("attn_weights_145_transpose_x_0"), val = bool(false)];
            bool attn_weights_145_transpose_y_0 = const()[name = string("attn_weights_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_145_cast_fp16 = matmul(transpose_x = attn_weights_145_transpose_x_0, transpose_y = attn_weights_145_transpose_y_0, x = var_3573_cast_fp16_0, y = var_3586_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3589_to_fp16 = const()[name = string("op_3589_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_147_cast_fp16 = mul(x = attn_weights_145_cast_fp16, y = var_3589_to_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_149_cast_fp16 = add(x = attn_weights_147_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_149_cast_fp16")];
            int32 var_3593 = const()[name = string("op_3593"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_151_cast_fp16 = softmax(axis = var_3593, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3599_transpose_x_1 = const()[name = string("op_3599_transpose_x_1"), val = bool(true)];
            bool var_3599_transpose_y_1 = const()[name = string("op_3599_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_3599_cast_fp16 = matmul(transpose_x = var_3599_transpose_x_1, transpose_y = var_3599_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3583_cast_fp16_0)[name = string("op_3599_cast_fp16")];
            bool attn_weights_153_transpose_x_0 = const()[name = string("attn_weights_153_transpose_x_0"), val = bool(false)];
            bool attn_weights_153_transpose_y_0 = const()[name = string("attn_weights_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_153_cast_fp16 = matmul(transpose_x = attn_weights_153_transpose_x_0, transpose_y = attn_weights_153_transpose_y_0, x = var_3573_cast_fp16_1, y = var_3586_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3601_to_fp16 = const()[name = string("op_3601_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_155_cast_fp16 = mul(x = attn_weights_153_cast_fp16, y = var_3601_to_fp16)[name = string("attn_weights_155_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_157_cast_fp16 = add(x = attn_weights_155_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_157_cast_fp16")];
            int32 var_3605 = const()[name = string("op_3605"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_159_cast_fp16 = softmax(axis = var_3605, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3583_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3613 = const()[name = string("op_3613"), val = int32(1)];
            bool attn_output_75_interleave_0 = const()[name = string("attn_output_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_75_cast_fp16 = concat(axis = var_3613, interleave = attn_output_75_interleave_0, values = (var_3599_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3617_perm_0 = const()[name = string("op_3617_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_3617_cast_fp16 = transpose(perm = var_3617_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_237")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3617_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_3650_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3650_cast_fp16")];
            int32 var_3648 = const()[name = string("op_3648"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_77_cast_fp16 = concat(axis = var_3648, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3650_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872055552)))];
            fp16 var_3660_to_fp16 = const()[name = string("op_3660_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3660_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3671_split_sizes_0 = const()[name = string("op_3671_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3671_axis_0 = const()[name = string("op_3671_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3671_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3671_cast_fp16_1 = split(axis = var_3671_axis_0, split_sizes = var_3671_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3671_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_3688_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3688_cast_fp16")];
            tensor<int32, [2]> var_3694_strides_0 = const()[name = string("op_3694_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3694_pad_type_0 = const()[name = string("op_3694_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3694_pad_0 = const()[name = string("op_3694_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3694_dilations_0 = const()[name = string("op_3694_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3694_groups_0 = const()[name = string("op_3694_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_3694_cast_fp16 = conv(dilations = var_3694_dilations_0, groups = var_3694_groups_0, pad = var_3694_pad_0, pad_type = var_3694_pad_type_0, strides = var_3694_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("op_3694_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_99_cast_fp16 = mul(x = var_3688_cast_fp16, y = var_3694_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_3712_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3712_cast_fp16")];
            int32 var_3710 = const()[name = string("op_3710"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_81_cast_fp16 = concat(axis = var_3710, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3712_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872063808)))];
            fp16 var_3722_to_fp16 = const()[name = string("op_3722_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3722_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3733_split_sizes_0 = const()[name = string("op_3733_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3733_axis_0 = const()[name = string("op_3733_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3733_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3733_cast_fp16_1 = split(axis = var_3733_axis_0, split_sizes = var_3733_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872072064)))];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(880460736)))];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_cast_fp16, x = var_3733_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_3790_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3790_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_3797_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3797_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3801_cast_fp16 = mul(x = x_101_cast_fp16, y = var_452_cast_fp16)[name = string("op_3801_cast_fp16")];
            tensor<int32, [2]> var_3802_split_sizes_0 = const()[name = string("op_3802_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3802_axis_0 = const()[name = string("op_3802_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_3802_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_3802_cast_fp16_1 = split(axis = var_3802_axis_0, split_sizes = var_3802_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3802_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_3804_cast_fp16 = mul(x = var_3802_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3804_cast_fp16")];
            int32 var_3806 = const()[name = string("op_3806"), val = int32(-2)];
            bool var_3807_interleave_0 = const()[name = string("op_3807_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_3807_cast_fp16 = concat(axis = var_3806, interleave = var_3807_interleave_0, values = (var_3804_cast_fp16, var_3802_cast_fp16_0))[name = string("op_3807_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_3808_cast_fp16 = mul(x = var_3807_cast_fp16, y = var_459_cast_fp16)[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_63_cast_fp16 = add(x = var_3801_cast_fp16, y = var_3808_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3814_cast_fp16 = mul(x = var_3790_cast_fp16, y = var_452_cast_fp16)[name = string("op_3814_cast_fp16")];
            tensor<int32, [2]> var_3815_split_sizes_0 = const()[name = string("op_3815_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3815_axis_0 = const()[name = string("op_3815_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_3815_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_3815_cast_fp16_1 = split(axis = var_3815_axis_0, split_sizes = var_3815_split_sizes_0, x = var_3790_cast_fp16)[name = string("op_3815_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_3817_cast_fp16 = mul(x = var_3815_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3817_cast_fp16")];
            int32 var_3819 = const()[name = string("op_3819"), val = int32(-2)];
            bool var_3820_interleave_0 = const()[name = string("op_3820_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3820_cast_fp16 = concat(axis = var_3819, interleave = var_3820_interleave_0, values = (var_3817_cast_fp16, var_3815_cast_fp16_0))[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3821_cast_fp16 = mul(x = var_3820_cast_fp16, y = var_459_cast_fp16)[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_105_cast_fp16 = add(x = var_3814_cast_fp16, y = var_3821_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_236")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_158)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_160_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_160 = read_state(input = key_cache)[name = string("coreml_update_state_160")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3797_cast_fp16)[name = string("transpose_235")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_159)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_161_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_161 = read_state(input = value_cache)[name = string("coreml_update_state_161")];
            tensor<int32, [4]> var_3891_begin_0 = const()[name = string("op_3891_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3891_end_0 = const()[name = string("op_3891_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3891_end_mask_0 = const()[name = string("op_3891_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3891_cast_fp16 = slice_by_index(begin = var_3891_begin_0, end = var_3891_end_0, end_mask = var_3891_end_mask_0, x = coreml_update_state_160)[name = string("op_3891_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3894_axis_0 = const()[name = string("op_3894_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_1 = split(axis = var_3894_axis_0, split_sizes = tile_20, x = var_3891_cast_fp16)[name = string("op_3894_cast_fp16")];
            tensor<int32, [4]> var_3901_begin_0 = const()[name = string("op_3901_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3901_end_0 = const()[name = string("op_3901_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3901_end_mask_0 = const()[name = string("op_3901_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3901_cast_fp16 = slice_by_index(begin = var_3901_begin_0, end = var_3901_end_0, end_mask = var_3901_end_mask_0, x = coreml_update_state_161)[name = string("op_3901_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3904_axis_0 = const()[name = string("op_3904_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_1 = split(axis = var_3904_axis_0, split_sizes = tile_21, x = var_3901_cast_fp16)[name = string("op_3904_cast_fp16")];
            tensor<int32, [2]> var_3907_split_sizes_0 = const()[name = string("op_3907_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3907_axis_0 = const()[name = string("op_3907_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_3907_0, tensor<fp16, [1, 8, 128, 64]> var_3907_1 = split(axis = var_3907_axis_0, split_sizes = var_3907_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3907")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3894_cast_fp16_0, y = var_3907_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3910_to_fp16 = const()[name = string("op_3910_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3910_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3914 = const()[name = string("op_3914"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_167_cast_fp16 = softmax(axis = var_3914, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3920_transpose_x_1 = const()[name = string("op_3920_transpose_x_1"), val = bool(true)];
            bool var_3920_transpose_y_1 = const()[name = string("op_3920_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_3920_cast_fp16 = matmul(transpose_x = var_3920_transpose_x_1, transpose_y = var_3920_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3904_cast_fp16_0)[name = string("op_3920_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3894_cast_fp16_1, y = var_3907_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3922_to_fp16 = const()[name = string("op_3922_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3922_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3926 = const()[name = string("op_3926"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_175_cast_fp16 = softmax(axis = var_3926, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3904_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3934 = const()[name = string("op_3934"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_83_cast_fp16 = concat(axis = var_3934, interleave = attn_output_83_interleave_0, values = (var_3920_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3938_perm_0 = const()[name = string("op_3938_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_3938_cast_fp16 = transpose(perm = var_3938_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_234")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3938_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_3971_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3971_cast_fp16")];
            int32 var_3969 = const()[name = string("op_3969"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_85_cast_fp16 = concat(axis = var_3969, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3971_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881509376)))];
            fp16 var_3981_to_fp16 = const()[name = string("op_3981_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3981_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3992_split_sizes_0 = const()[name = string("op_3992_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3992_axis_0 = const()[name = string("op_3992_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_3992_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_3992_cast_fp16_1 = split(axis = var_3992_axis_0, split_sizes = var_3992_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3992_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_4009_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [2]> var_4015_strides_0 = const()[name = string("op_4015_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4015_pad_type_0 = const()[name = string("op_4015_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4015_pad_0 = const()[name = string("op_4015_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4015_dilations_0 = const()[name = string("op_4015_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4015_groups_0 = const()[name = string("op_4015_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_4015_cast_fp16 = conv(dilations = var_4015_dilations_0, groups = var_4015_groups_0, pad = var_4015_pad_0, pad_type = var_4015_pad_type_0, strides = var_4015_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("op_4015_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_109_cast_fp16 = mul(x = var_4009_cast_fp16, y = var_4015_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_4033_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4033_cast_fp16")];
            int32 var_4031 = const()[name = string("op_4031"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_89_cast_fp16 = concat(axis = var_4031, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4033_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881517632)))];
            fp16 var_4043_to_fp16 = const()[name = string("op_4043_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4043_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4054_split_sizes_0 = const()[name = string("op_4054_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4054_axis_0 = const()[name = string("op_4054_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_4054_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_4054_cast_fp16_1 = split(axis = var_4054_axis_0, split_sizes = var_4054_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_4111_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4111_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_4118_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4118_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_4122_cast_fp16 = mul(x = x_111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4122_cast_fp16")];
            tensor<int32, [2]> var_4123_split_sizes_0 = const()[name = string("op_4123_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4123_axis_0 = const()[name = string("op_4123_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_4123_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_4123_cast_fp16_1 = split(axis = var_4123_axis_0, split_sizes = var_4123_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4123_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_4125_cast_fp16 = mul(x = var_4123_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4125_cast_fp16")];
            int32 var_4127 = const()[name = string("op_4127"), val = int32(-2)];
            bool var_4128_interleave_0 = const()[name = string("op_4128_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_4128_cast_fp16 = concat(axis = var_4127, interleave = var_4128_interleave_0, values = (var_4125_cast_fp16, var_4123_cast_fp16_0))[name = string("op_4128_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_4129_cast_fp16 = mul(x = var_4128_cast_fp16, y = var_459_cast_fp16)[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_69_cast_fp16 = add(x = var_4122_cast_fp16, y = var_4129_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_4135_cast_fp16 = mul(x = var_4111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4135_cast_fp16")];
            tensor<int32, [2]> var_4136_split_sizes_0 = const()[name = string("op_4136_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4136_axis_0 = const()[name = string("op_4136_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_4136_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_4136_cast_fp16_1 = split(axis = var_4136_axis_0, split_sizes = var_4136_split_sizes_0, x = var_4111_cast_fp16)[name = string("op_4136_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_4138_cast_fp16 = mul(x = var_4136_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4138_cast_fp16")];
            int32 var_4140 = const()[name = string("op_4140"), val = int32(-2)];
            bool var_4141_interleave_0 = const()[name = string("op_4141_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_4141_cast_fp16 = concat(axis = var_4140, interleave = var_4141_interleave_0, values = (var_4138_cast_fp16, var_4136_cast_fp16_0))[name = string("op_4141_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_4142_cast_fp16 = mul(x = var_4141_cast_fp16, y = var_459_cast_fp16)[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_115_cast_fp16 = add(x = var_4135_cast_fp16, y = var_4142_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_233")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_160)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_162_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_162 = read_state(input = key_cache)[name = string("coreml_update_state_162")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4118_cast_fp16)[name = string("transpose_232")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_161)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_163_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_163 = read_state(input = value_cache)[name = string("coreml_update_state_163")];
            tensor<int32, [4]> var_4212_begin_0 = const()[name = string("op_4212_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4212_end_0 = const()[name = string("op_4212_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4212_end_mask_0 = const()[name = string("op_4212_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4212_cast_fp16 = slice_by_index(begin = var_4212_begin_0, end = var_4212_end_0, end_mask = var_4212_end_mask_0, x = coreml_update_state_162)[name = string("op_4212_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4215_axis_0 = const()[name = string("op_4215_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_1 = split(axis = var_4215_axis_0, split_sizes = tile_22, x = var_4212_cast_fp16)[name = string("op_4215_cast_fp16")];
            tensor<int32, [4]> var_4222_begin_0 = const()[name = string("op_4222_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4222_end_0 = const()[name = string("op_4222_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4222_end_mask_0 = const()[name = string("op_4222_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4222_cast_fp16 = slice_by_index(begin = var_4222_begin_0, end = var_4222_end_0, end_mask = var_4222_end_mask_0, x = coreml_update_state_163)[name = string("op_4222_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4225_axis_0 = const()[name = string("op_4225_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_1 = split(axis = var_4225_axis_0, split_sizes = tile_23, x = var_4222_cast_fp16)[name = string("op_4225_cast_fp16")];
            tensor<int32, [2]> var_4228_split_sizes_0 = const()[name = string("op_4228_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4228_axis_0 = const()[name = string("op_4228_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_4228_0, tensor<fp16, [1, 8, 128, 64]> var_4228_1 = split(axis = var_4228_axis_0, split_sizes = var_4228_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4228")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4215_cast_fp16_0, y = var_4228_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4231_to_fp16 = const()[name = string("op_4231_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4231_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4235 = const()[name = string("op_4235"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_183_cast_fp16 = softmax(axis = var_4235, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4241_transpose_x_1 = const()[name = string("op_4241_transpose_x_1"), val = bool(true)];
            bool var_4241_transpose_y_1 = const()[name = string("op_4241_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_4241_cast_fp16 = matmul(transpose_x = var_4241_transpose_x_1, transpose_y = var_4241_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4225_cast_fp16_0)[name = string("op_4241_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4215_cast_fp16_1, y = var_4228_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4243_to_fp16 = const()[name = string("op_4243_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4243_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4247 = const()[name = string("op_4247"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_191_cast_fp16 = softmax(axis = var_4247, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4225_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4255 = const()[name = string("op_4255"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_91_cast_fp16 = concat(axis = var_4255, interleave = attn_output_91_interleave_0, values = (var_4241_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4259_perm_0 = const()[name = string("op_4259_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_4259_cast_fp16 = transpose(perm = var_4259_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_231")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4259_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_4292_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4292_cast_fp16")];
            int32 var_4290 = const()[name = string("op_4290"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_93_cast_fp16 = concat(axis = var_4290, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4292_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881525888)))];
            fp16 var_4302_to_fp16 = const()[name = string("op_4302_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4302_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4313_split_sizes_0 = const()[name = string("op_4313_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4313_axis_0 = const()[name = string("op_4313_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_4313_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_4313_cast_fp16_1 = split(axis = var_4313_axis_0, split_sizes = var_4313_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4313_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_4330_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4330_cast_fp16")];
            tensor<int32, [2]> var_4336_strides_0 = const()[name = string("op_4336_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4336_pad_type_0 = const()[name = string("op_4336_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4336_pad_0 = const()[name = string("op_4336_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4336_dilations_0 = const()[name = string("op_4336_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4336_groups_0 = const()[name = string("op_4336_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_4336_cast_fp16 = conv(dilations = var_4336_dilations_0, groups = var_4336_groups_0, pad = var_4336_pad_0, pad_type = var_4336_pad_type_0, strides = var_4336_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("op_4336_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_119_cast_fp16 = mul(x = var_4330_cast_fp16, y = var_4336_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_cast_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_4354_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4354_cast_fp16")];
            int32 var_4352 = const()[name = string("op_4352"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_97_cast_fp16 = concat(axis = var_4352, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4354_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881534144)))];
            fp16 var_4364_to_fp16 = const()[name = string("op_4364_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4364_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4375_split_sizes_0 = const()[name = string("op_4375_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4375_axis_0 = const()[name = string("op_4375_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_4375_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_4375_cast_fp16_1 = split(axis = var_4375_axis_0, split_sizes = var_4375_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4375_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_4432_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4432_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_4439_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4439_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_4443_cast_fp16 = mul(x = x_121_cast_fp16, y = var_452_cast_fp16)[name = string("op_4443_cast_fp16")];
            tensor<int32, [2]> var_4444_split_sizes_0 = const()[name = string("op_4444_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4444_axis_0 = const()[name = string("op_4444_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_4444_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_4444_cast_fp16_1 = split(axis = var_4444_axis_0, split_sizes = var_4444_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4444_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_4446_cast_fp16 = mul(x = var_4444_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4446_cast_fp16")];
            int32 var_4448 = const()[name = string("op_4448"), val = int32(-2)];
            bool var_4449_interleave_0 = const()[name = string("op_4449_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_4449_cast_fp16 = concat(axis = var_4448, interleave = var_4449_interleave_0, values = (var_4446_cast_fp16, var_4444_cast_fp16_0))[name = string("op_4449_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_4450_cast_fp16 = mul(x = var_4449_cast_fp16, y = var_459_cast_fp16)[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_75_cast_fp16 = add(x = var_4443_cast_fp16, y = var_4450_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_4456_cast_fp16 = mul(x = var_4432_cast_fp16, y = var_452_cast_fp16)[name = string("op_4456_cast_fp16")];
            tensor<int32, [2]> var_4457_split_sizes_0 = const()[name = string("op_4457_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4457_axis_0 = const()[name = string("op_4457_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_4457_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_4457_cast_fp16_1 = split(axis = var_4457_axis_0, split_sizes = var_4457_split_sizes_0, x = var_4432_cast_fp16)[name = string("op_4457_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_4459_cast_fp16 = mul(x = var_4457_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4459_cast_fp16")];
            int32 var_4461 = const()[name = string("op_4461"), val = int32(-2)];
            bool var_4462_interleave_0 = const()[name = string("op_4462_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_4462_cast_fp16 = concat(axis = var_4461, interleave = var_4462_interleave_0, values = (var_4459_cast_fp16, var_4457_cast_fp16_0))[name = string("op_4462_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_4463_cast_fp16 = mul(x = var_4462_cast_fp16, y = var_459_cast_fp16)[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_125_cast_fp16 = add(x = var_4456_cast_fp16, y = var_4463_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_230")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_162)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_164_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_164 = read_state(input = key_cache)[name = string("coreml_update_state_164")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4439_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_163)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_165_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_165 = read_state(input = value_cache)[name = string("coreml_update_state_165")];
            tensor<int32, [4]> var_4533_begin_0 = const()[name = string("op_4533_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4533_end_0 = const()[name = string("op_4533_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4533_end_mask_0 = const()[name = string("op_4533_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4533_cast_fp16 = slice_by_index(begin = var_4533_begin_0, end = var_4533_end_0, end_mask = var_4533_end_mask_0, x = coreml_update_state_164)[name = string("op_4533_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4536_axis_0 = const()[name = string("op_4536_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_1 = split(axis = var_4536_axis_0, split_sizes = tile_24, x = var_4533_cast_fp16)[name = string("op_4536_cast_fp16")];
            tensor<int32, [4]> var_4543_begin_0 = const()[name = string("op_4543_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4543_end_0 = const()[name = string("op_4543_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4543_end_mask_0 = const()[name = string("op_4543_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4543_cast_fp16 = slice_by_index(begin = var_4543_begin_0, end = var_4543_end_0, end_mask = var_4543_end_mask_0, x = coreml_update_state_165)[name = string("op_4543_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4546_axis_0 = const()[name = string("op_4546_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_1 = split(axis = var_4546_axis_0, split_sizes = tile_25, x = var_4543_cast_fp16)[name = string("op_4546_cast_fp16")];
            tensor<int32, [2]> var_4549_split_sizes_0 = const()[name = string("op_4549_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4549_axis_0 = const()[name = string("op_4549_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_4549_0, tensor<fp16, [1, 8, 128, 64]> var_4549_1 = split(axis = var_4549_axis_0, split_sizes = var_4549_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4549")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4536_cast_fp16_0, y = var_4549_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4552_to_fp16 = const()[name = string("op_4552_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4552_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4556 = const()[name = string("op_4556"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_199_cast_fp16 = softmax(axis = var_4556, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4562_transpose_x_1 = const()[name = string("op_4562_transpose_x_1"), val = bool(true)];
            bool var_4562_transpose_y_1 = const()[name = string("op_4562_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_4562_cast_fp16 = matmul(transpose_x = var_4562_transpose_x_1, transpose_y = var_4562_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4546_cast_fp16_0)[name = string("op_4562_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4536_cast_fp16_1, y = var_4549_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4564_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4568 = const()[name = string("op_4568"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_207_cast_fp16 = softmax(axis = var_4568, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4546_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4576 = const()[name = string("op_4576"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_99_cast_fp16 = concat(axis = var_4576, interleave = attn_output_99_interleave_0, values = (var_4562_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4580_perm_0 = const()[name = string("op_4580_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_4580_cast_fp16 = transpose(perm = var_4580_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_228")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4580_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_4613_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4613_cast_fp16")];
            int32 var_4611 = const()[name = string("op_4611"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_101_cast_fp16 = concat(axis = var_4611, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4613_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881542400)))];
            fp16 var_4623_to_fp16 = const()[name = string("op_4623_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4623_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4634_split_sizes_0 = const()[name = string("op_4634_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4634_axis_0 = const()[name = string("op_4634_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_4634_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_4634_cast_fp16_1 = split(axis = var_4634_axis_0, split_sizes = var_4634_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4634_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_4651_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4651_cast_fp16")];
            tensor<int32, [2]> var_4657_strides_0 = const()[name = string("op_4657_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4657_pad_type_0 = const()[name = string("op_4657_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4657_pad_0 = const()[name = string("op_4657_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4657_dilations_0 = const()[name = string("op_4657_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4657_groups_0 = const()[name = string("op_4657_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_4657_cast_fp16 = conv(dilations = var_4657_dilations_0, groups = var_4657_groups_0, pad = var_4657_pad_0, pad_type = var_4657_pad_type_0, strides = var_4657_strides_0, weight = layers_12_mlp_up_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("op_4657_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_129_cast_fp16 = mul(x = var_4651_cast_fp16, y = var_4657_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_cast_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_4675_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4675_cast_fp16")];
            int32 var_4673 = const()[name = string("op_4673"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_105_cast_fp16 = concat(axis = var_4673, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4675_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881550656)))];
            fp16 var_4685_to_fp16 = const()[name = string("op_4685_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4685_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4696_split_sizes_0 = const()[name = string("op_4696_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4696_axis_0 = const()[name = string("op_4696_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_4696_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_4696_cast_fp16_1 = split(axis = var_4696_axis_0, split_sizes = var_4696_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4696_cast_fp16")];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 64]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 64]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_4753_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4753_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 64]> var_4760_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4760_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_4764_cast_fp16 = mul(x = x_131_cast_fp16, y = var_452_cast_fp16)[name = string("op_4764_cast_fp16")];
            tensor<int32, [2]> var_4765_split_sizes_0 = const()[name = string("op_4765_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4765_axis_0 = const()[name = string("op_4765_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 64]> var_4765_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_4765_cast_fp16_1 = split(axis = var_4765_axis_0, split_sizes = var_4765_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4765_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_4767_cast_fp16 = mul(x = var_4765_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4767_cast_fp16")];
            int32 var_4769 = const()[name = string("op_4769"), val = int32(-2)];
            bool var_4770_interleave_0 = const()[name = string("op_4770_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 64]> var_4770_cast_fp16 = concat(axis = var_4769, interleave = var_4770_interleave_0, values = (var_4767_cast_fp16, var_4765_cast_fp16_0))[name = string("op_4770_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_4771_cast_fp16 = mul(x = var_4770_cast_fp16, y = var_459_cast_fp16)[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> query_states_81_cast_fp16 = add(x = var_4764_cast_fp16, y = var_4771_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_4777_cast_fp16 = mul(x = var_4753_cast_fp16, y = var_452_cast_fp16)[name = string("op_4777_cast_fp16")];
            tensor<int32, [2]> var_4778_split_sizes_0 = const()[name = string("op_4778_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4778_axis_0 = const()[name = string("op_4778_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 64]> var_4778_cast_fp16_0, tensor<fp16, [1, 2, 64, 64]> var_4778_cast_fp16_1 = split(axis = var_4778_axis_0, split_sizes = var_4778_split_sizes_0, x = var_4753_cast_fp16)[name = string("op_4778_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 64]> var_4780_cast_fp16 = mul(x = var_4778_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4780_cast_fp16")];
            int32 var_4782 = const()[name = string("op_4782"), val = int32(-2)];
            bool var_4783_interleave_0 = const()[name = string("op_4783_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_4783_cast_fp16 = concat(axis = var_4782, interleave = var_4783_interleave_0, values = (var_4780_cast_fp16, var_4778_cast_fp16_0))[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_4784_cast_fp16 = mul(x = var_4783_cast_fp16, y = var_459_cast_fp16)[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> key_states_135_cast_fp16 = add(x = var_4777_cast_fp16, y = var_4784_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_227")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_164)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_166_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_166 = read_state(input = key_cache)[name = string("coreml_update_state_166")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 64, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4760_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_165)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_167_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_167 = read_state(input = value_cache)[name = string("coreml_update_state_167")];
            tensor<int32, [4]> var_4854_begin_0 = const()[name = string("op_4854_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4854_end_0 = const()[name = string("op_4854_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4854_end_mask_0 = const()[name = string("op_4854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4854_cast_fp16 = slice_by_index(begin = var_4854_begin_0, end = var_4854_end_0, end_mask = var_4854_end_mask_0, x = coreml_update_state_166)[name = string("op_4854_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4857_axis_0 = const()[name = string("op_4857_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_1 = split(axis = var_4857_axis_0, split_sizes = tile_26, x = var_4854_cast_fp16)[name = string("op_4857_cast_fp16")];
            tensor<int32, [4]> var_4864_begin_0 = const()[name = string("op_4864_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4864_end_0 = const()[name = string("op_4864_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4864_end_mask_0 = const()[name = string("op_4864_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4864_cast_fp16 = slice_by_index(begin = var_4864_begin_0, end = var_4864_end_0, end_mask = var_4864_end_mask_0, x = coreml_update_state_167)[name = string("op_4864_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4867_axis_0 = const()[name = string("op_4867_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_1 = split(axis = var_4867_axis_0, split_sizes = tile_27, x = var_4864_cast_fp16)[name = string("op_4867_cast_fp16")];
            tensor<int32, [2]> var_4870_split_sizes_0 = const()[name = string("op_4870_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4870_axis_0 = const()[name = string("op_4870_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 64]> var_4870_0, tensor<fp16, [1, 8, 128, 64]> var_4870_1 = split(axis = var_4870_axis_0, split_sizes = var_4870_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4870")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4857_cast_fp16_0, y = var_4870_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4873_to_fp16 = const()[name = string("op_4873_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4873_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4877 = const()[name = string("op_4877"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_215_cast_fp16 = softmax(axis = var_4877, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4883_transpose_x_1 = const()[name = string("op_4883_transpose_x_1"), val = bool(true)];
            bool var_4883_transpose_y_1 = const()[name = string("op_4883_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_4883_cast_fp16 = matmul(transpose_x = var_4883_transpose_x_1, transpose_y = var_4883_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4867_cast_fp16_0)[name = string("op_4883_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4857_cast_fp16_1, y = var_4870_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4885_to_fp16 = const()[name = string("op_4885_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4885_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4889 = const()[name = string("op_4889"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 64]> attn_weights_cast_fp16 = softmax(axis = var_4889, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4867_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4897 = const()[name = string("op_4897"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> attn_output_107_cast_fp16 = concat(axis = var_4897, interleave = attn_output_107_interleave_0, values = (var_4883_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4901_perm_0 = const()[name = string("op_4901_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 64]> var_4901_cast_fp16 = transpose(perm = var_4901_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_225")];
            tensor<fp16, [1, 2048, 1, 64]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4901_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 64]> var_4934_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4934_cast_fp16")];
            int32 var_4932 = const()[name = string("op_4932"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 64]> doubled_109_cast_fp16 = concat(axis = var_4932, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4934_cast_fp16))[name = string("doubled_109_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(881558912)))];
            fp16 var_4944_to_fp16 = const()[name = string("op_4944_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 64]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_4944_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_4955_split_sizes_0 = const()[name = string("op_4955_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4955_axis_0 = const()[name = string("op_4955_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> var_4955_cast_fp16_0, tensor<fp16, [1, 2048, 1, 64]> var_4955_cast_fp16_1 = split(axis = var_4955_axis_0, split_sizes = var_4955_split_sizes_0, x = out_cast_fp16)[name = string("op_4955_cast_fp16")];
            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_13_mlp_gate_proj_weight_cast_fp16, x = var_4955_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> var_4972_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4972_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881567168)))];
            tensor<int32, [2]> var_4978_strides_0 = const()[name = string("op_4978_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4978_pad_type_0 = const()[name = string("op_4978_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4978_pad_0 = const()[name = string("op_4978_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4978_dilations_0 = const()[name = string("op_4978_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4978_groups_0 = const()[name = string("op_4978_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 64]> var_4978_cast_fp16 = conv(dilations = var_4978_dilations_0, groups = var_4978_groups_0, pad = var_4978_pad_0, pad_type = var_4978_pad_type_0, strides = var_4978_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4955_cast_fp16_0)[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 64]> x_cast_fp16 = mul(x = var_4972_cast_fp16, y = var_4978_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_13_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 64]> hidden_states = add(x = hidden_states_135_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_4987_cast_fp16")];
        } -> (hidden_states);
    func length_8<ios18>(tensor<fp32, [1, 2048, 1, 8]> inputs_embeds, state<tensor<fp16, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [8]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_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(64))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524992))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524416))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_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(525312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13120640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13108288))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_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(13126848))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651776))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651200))))[name = string("layers_2_self_attn_v_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(13652096))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26247424))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26235072))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_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(26253632))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26778560))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26777984))))[name = string("layers_3_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_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(26778880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30977408))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30973248))))[name = string("layers_3_self_attn_o_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(30979520))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43566656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43562496))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_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(43568768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093120))))[name = string("layers_4_self_attn_v_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(44094016))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48292544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48288384))))[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(48294656))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60889984))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60877632))))[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(60896192))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73491520))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73479168))))[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(73497728))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86084864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86080704))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_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(86086976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611328))))[name = string("layers_5_self_attn_v_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(86612224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90810752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90806592))))[name = string("layers_5_self_attn_o_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(90812864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103395840))))[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(103414400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997376))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_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(116003648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528000))))[name = string("layers_6_self_attn_v_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(116528896))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120727424))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120723264))))[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(120729536))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133324864))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133312512))))[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(133331072))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145926400))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145914048))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_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(145932608))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158519744))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158515584))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_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(158521856))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046784))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046208))))[name = string("layers_7_self_attn_v_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(159047104))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163245632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241472))))[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(163247744))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175843072))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175830720))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_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(175849280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176374208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176373632))))[name = string("layers_8_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_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(176374528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180573056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180568896))))[name = string("layers_8_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(180575168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193170496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193158144))))[name = string("layers_8_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(193176704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205772032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205759680))))[name = string("layers_8_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_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(205778240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218365376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218361216))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_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(218367488))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218892416))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218891840))))[name = string("layers_9_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_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(218892736))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223091264))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223087104))))[name = string("layers_9_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(223093376))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235688704))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235676352))))[name = string("layers_9_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(235694912))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248290240))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248277888))))[name = string("layers_9_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_9_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(248296448))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260883584))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260879424))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_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(260885696))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410624))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410048))))[name = string("layers_10_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_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(261410944))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265609472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265605312))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(265611584))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278206912))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278194560))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(278213120))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290808448))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290796096))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_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(290814656))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303401792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303397632))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(303403904))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307602432))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307598272))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_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(307604544))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308129472))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308128896))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_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(308129792))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654720))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654144))))[name = string("layers_11_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(308655040))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312853568))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312849408))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(312855680))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325451008))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325438656))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(325457216))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338052544))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040192))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_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(338058752))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350645888))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350641728))))[name = string("layers_11_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(350648000))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354846528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354842368))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_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(354848640))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355373568))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355372992))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_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(355373888))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898816))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898240))))[name = string("layers_12_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(355899136))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360097664))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360093504))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(360099776))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372695104))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372682752))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(372701312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385296640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385284288))))[name = string("layers_12_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_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(385302848))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397889984))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397885824))))[name = string("layers_12_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(397892096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402090624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402086464))))[name = string("layers_13_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_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(402092736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617088))))[name = string("layers_13_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_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(402617984))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142912))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142336))))[name = string("layers_13_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(403143232))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407341760))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407337600))))[name = string("layers_13_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_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(407343872))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419939200))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419926848))))[name = string("layers_13_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_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(419945408))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432532544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432528384))))[name = string("layers_13_mlp_down_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_424 = const()[name = string("op_424"), val = int32(0)];
            bool var_426_exclusive_0 = const()[name = string("op_426_exclusive_0"), val = bool(false)];
            bool var_426_reverse_0 = const()[name = string("op_426_reverse_0"), val = bool(false)];
            tensor<fp16, [8]> var_426_cast_fp16 = cumsum(axis = var_424, exclusive = var_426_exclusive_0, reverse = var_426_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_426_cast_fp16")];
            fp16 var_428_promoted_to_fp16 = const()[name = string("op_428_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [8]> position_offsets_cast_fp16 = sub(x = var_426_cast_fp16, y = var_428_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_431_axes_0 = const()[name = string("op_431_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8]> var_431_cast_fp16 = expand_dims(axes = var_431_axes_0, x = position_offsets_cast_fp16)[name = string("op_431_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_431_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(432534656)))];
            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(440923328)))];
            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_450_perm_0 = const()[name = string("op_450_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_452_axes_0 = const()[name = string("op_452_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 8]> var_450_cast_fp16 = transpose(perm = var_450_perm_0, x = cos_1_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 1, 128, 8]> var_452_cast_fp16 = expand_dims(axes = var_452_axes_0, x = var_450_cast_fp16)[name = string("op_452_cast_fp16")];
            tensor<int32, [3]> var_457_perm_0 = const()[name = string("op_457_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_459_axes_0 = const()[name = string("op_459_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 8]> var_457_cast_fp16 = transpose(perm = var_457_perm_0, x = sin_1_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 1, 128, 8]> var_459_cast_fp16 = expand_dims(axes = var_459_axes_0, x = var_457_cast_fp16)[name = string("op_459_cast_fp16")];
            tensor<int32, [1]> var_478_axes_0 = const()[name = string("op_478_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 8, 1]> var_478 = expand_dims(axes = var_478_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_478")];
            tensor<int32, [1, 1, 2048]> var_471 = const()[name = string("op_471"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449312000)))];
            tensor<bool, [1, 8, 2048]> var_479 = greater(x = var_471, y = var_478)[name = string("op_479")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_486_axes_0 = const()[name = string("op_486_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 8, 2048]> var_479_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_479)[name = string("cast_9")];
            tensor<fp16, [1, 1, 8, 2048]> var_486_cast_fp16 = expand_dims(axes = var_486_axes_0, x = var_479_to_fp16)[name = string("op_486_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_490_promoted_to_fp16 = const()[name = string("op_490_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 8]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_486_cast_fp16)[name = string("transpose_132")];
            tensor<bool, [1, 1, 2048, 8]> var_491_cast_fp16 = equal(x = mask_cast_fp16, y = var_490_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            fp16 var_492_to_fp16 = const()[name = string("op_492_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 8]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_492_to_fp16, cond = var_491_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_502_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_502_cast_fp16")];
            int32 var_500 = const()[name = string("op_500"), 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_500, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_502_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(449320256)))];
            fp16 var_512_to_fp16 = const()[name = string("op_512_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_512_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_523_split_sizes_0 = const()[name = string("op_523_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_523_axis_0 = const()[name = string("op_523_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_523_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_523_cast_fp16_1 = split(axis = var_523_axis_0, split_sizes = var_523_split_sizes_0, x = out_1_cast_fp16)[name = string("op_523_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449328512)))];
            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_to_fp16, x = var_523_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457717184)))];
            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_to_fp16, x = var_523_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(458765824)))];
            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_523_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_580_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_580_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_587_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_587_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_591_cast_fp16 = mul(x = x_1_cast_fp16, y = var_452_cast_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [2]> var_592_split_sizes_0 = const()[name = string("op_592_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_592_axis_0 = const()[name = string("op_592_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_592_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_592_cast_fp16_1 = split(axis = var_592_axis_0, split_sizes = var_592_split_sizes_0, x = x_1_cast_fp16)[name = string("op_592_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_594_cast_fp16 = mul(x = var_592_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_594_cast_fp16")];
            int32 var_596 = const()[name = string("op_596"), val = int32(-2)];
            bool var_597_interleave_0 = const()[name = string("op_597_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_597_cast_fp16 = concat(axis = var_596, interleave = var_597_interleave_0, values = (var_594_cast_fp16, var_592_cast_fp16_0))[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_598_cast_fp16 = mul(x = var_597_cast_fp16, y = var_459_cast_fp16)[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_3_cast_fp16 = add(x = var_591_cast_fp16, y = var_598_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_604_cast_fp16 = mul(x = var_580_cast_fp16, y = var_452_cast_fp16)[name = string("op_604_cast_fp16")];
            tensor<int32, [2]> var_605_split_sizes_0 = const()[name = string("op_605_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_605_axis_0 = const()[name = string("op_605_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_605_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_605_cast_fp16_1 = split(axis = var_605_axis_0, split_sizes = var_605_split_sizes_0, x = var_580_cast_fp16)[name = string("op_605_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_607_cast_fp16 = mul(x = var_605_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_607_cast_fp16")];
            int32 var_609 = const()[name = string("op_609"), val = int32(-2)];
            bool var_610_interleave_0 = const()[name = string("op_610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_610_cast_fp16 = concat(axis = var_609, interleave = var_610_interleave_0, values = (var_607_cast_fp16, var_605_cast_fp16_0))[name = string("op_610_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_611_cast_fp16 = mul(x = var_610_cast_fp16, y = var_459_cast_fp16)[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_5_cast_fp16 = add(x = var_604_cast_fp16, y = var_611_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 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_131")];
            tensor<fp16, [14, 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_56_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_56 = read_state(input = key_cache)[name = string("coreml_update_state_56")];
            tensor<fp16, [14, 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_587_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [14, 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_57_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_57 = read_state(input = value_cache)[name = string("coreml_update_state_57")];
            tensor<int32, [4]> var_681_begin_0 = const()[name = string("op_681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_681_end_0 = const()[name = string("op_681_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_681_end_mask_0 = const()[name = string("op_681_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_681_cast_fp16 = slice_by_index(begin = var_681_begin_0, end = var_681_end_0, end_mask = var_681_end_mask_0, x = coreml_update_state_56)[name = string("op_681_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_684_axis_0 = const()[name = string("op_684_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_1 = split(axis = var_684_axis_0, split_sizes = tile_0, x = var_681_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [4]> var_691_begin_0 = const()[name = string("op_691_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_691_end_0 = const()[name = string("op_691_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_691_end_mask_0 = const()[name = string("op_691_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = coreml_update_state_57)[name = string("op_691_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_694_axis_0 = const()[name = string("op_694_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_1 = split(axis = var_694_axis_0, split_sizes = tile_1, x = var_691_cast_fp16)[name = string("op_694_cast_fp16")];
            tensor<int32, [2]> var_697_split_sizes_0 = const()[name = string("op_697_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_697_axis_0 = const()[name = string("op_697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_697_0, tensor<fp16, [1, 8, 128, 8]> var_697_1 = split(axis = var_697_axis_0, split_sizes = var_697_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_697")];
            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_684_cast_fp16_0, y = var_697_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_700_to_fp16 = const()[name = string("op_700_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_700_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_704 = const()[name = string("op_704"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_7_cast_fp16 = softmax(axis = var_704, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_710_transpose_x_1 = const()[name = string("op_710_transpose_x_1"), val = bool(true)];
            bool var_710_transpose_y_1 = const()[name = string("op_710_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_710_cast_fp16 = matmul(transpose_x = var_710_transpose_x_1, transpose_y = var_710_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_694_cast_fp16_0)[name = string("op_710_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_684_cast_fp16_1, y = var_697_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_712_to_fp16 = const()[name = string("op_712_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_712_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_716 = const()[name = string("op_716"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_15_cast_fp16 = softmax(axis = var_716, 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_694_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_724 = const()[name = string("op_724"), 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_724, interleave = attn_output_3_interleave_0, values = (var_710_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_728_perm_0 = const()[name = string("op_728_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_728_cast_fp16 = transpose(perm = var_728_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_728_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(459814464)))];
            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_761_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_761_cast_fp16")];
            int32 var_759 = const()[name = string("op_759"), 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_759, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_761_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(468203136)))];
            fp16 var_771_to_fp16 = const()[name = string("op_771_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_771_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_782_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = out_3_cast_fp16)[name = string("op_782_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468211392)))];
            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_to_fp16, x = var_782_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_799_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_799_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493377280)))];
            tensor<int32, [2]> var_805_strides_0 = const()[name = string("op_805_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_805_pad_type_0 = const()[name = string("op_805_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_805_pad_0 = const()[name = string("op_805_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_805_dilations_0 = const()[name = string("op_805_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_805_groups_0 = const()[name = string("op_805_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_805_cast_fp16 = conv(dilations = var_805_dilations_0, groups = var_805_groups_0, pad = var_805_pad_0, pad_type = var_805_pad_type_0, strides = var_805_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_782_cast_fp16_0)[name = string("op_805_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_9_cast_fp16 = mul(x = var_799_cast_fp16, y = var_805_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518543168)))];
            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_to_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_823_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_823_cast_fp16")];
            int32 var_821 = const()[name = string("op_821"), 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_821, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_823_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(543709056)))];
            fp16 var_833_to_fp16 = const()[name = string("op_833_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_833_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_844_split_sizes_0 = const()[name = string("op_844_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_844_axis_0 = const()[name = string("op_844_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_844_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_844_cast_fp16_1 = split(axis = var_844_axis_0, split_sizes = var_844_split_sizes_0, x = out_5_cast_fp16)[name = string("op_844_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543717312)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552105984)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            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_cast_fp16, x = var_844_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_901_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_901_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_908_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_908_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_912_cast_fp16 = mul(x = x_11_cast_fp16, y = var_452_cast_fp16)[name = string("op_912_cast_fp16")];
            tensor<int32, [2]> var_913_split_sizes_0 = const()[name = string("op_913_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_913_axis_0 = const()[name = string("op_913_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_913_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_913_cast_fp16_1 = split(axis = var_913_axis_0, split_sizes = var_913_split_sizes_0, x = x_11_cast_fp16)[name = string("op_913_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_915_cast_fp16 = mul(x = var_913_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_915_cast_fp16")];
            int32 var_917 = const()[name = string("op_917"), val = int32(-2)];
            bool var_918_interleave_0 = const()[name = string("op_918_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_918_cast_fp16 = concat(axis = var_917, interleave = var_918_interleave_0, values = (var_915_cast_fp16, var_913_cast_fp16_0))[name = string("op_918_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_919_cast_fp16 = mul(x = var_918_cast_fp16, y = var_459_cast_fp16)[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_9_cast_fp16 = add(x = var_912_cast_fp16, y = var_919_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_925_cast_fp16 = mul(x = var_901_cast_fp16, y = var_452_cast_fp16)[name = string("op_925_cast_fp16")];
            tensor<int32, [2]> var_926_split_sizes_0 = const()[name = string("op_926_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_926_axis_0 = const()[name = string("op_926_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_926_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_926_cast_fp16_1 = split(axis = var_926_axis_0, split_sizes = var_926_split_sizes_0, x = var_901_cast_fp16)[name = string("op_926_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_928_cast_fp16 = mul(x = var_926_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_928_cast_fp16")];
            int32 var_930 = const()[name = string("op_930"), val = int32(-2)];
            bool var_931_interleave_0 = const()[name = string("op_931_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_931_cast_fp16 = concat(axis = var_930, interleave = var_931_interleave_0, values = (var_928_cast_fp16, var_926_cast_fp16_0))[name = string("op_931_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_932_cast_fp16 = mul(x = var_931_cast_fp16, y = var_459_cast_fp16)[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_15_cast_fp16 = add(x = var_925_cast_fp16, y = var_932_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_128")];
            tensor<fp16, [14, 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_56)[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_58_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_58 = read_state(input = key_cache)[name = string("coreml_update_state_58")];
            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_908_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [14, 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_57)[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_59_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_59 = read_state(input = value_cache)[name = string("coreml_update_state_59")];
            tensor<int32, [4]> var_1002_begin_0 = const()[name = string("op_1002_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1002_end_0 = const()[name = string("op_1002_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1002_end_mask_0 = const()[name = string("op_1002_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1002_cast_fp16 = slice_by_index(begin = var_1002_begin_0, end = var_1002_end_0, end_mask = var_1002_end_mask_0, x = coreml_update_state_58)[name = string("op_1002_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1005_axis_0 = const()[name = string("op_1005_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_1 = split(axis = var_1005_axis_0, split_sizes = tile_2, x = var_1002_cast_fp16)[name = string("op_1005_cast_fp16")];
            tensor<int32, [4]> var_1012_begin_0 = const()[name = string("op_1012_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1012_end_0 = const()[name = string("op_1012_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1012_end_mask_0 = const()[name = string("op_1012_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1012_cast_fp16 = slice_by_index(begin = var_1012_begin_0, end = var_1012_end_0, end_mask = var_1012_end_mask_0, x = coreml_update_state_59)[name = string("op_1012_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1015_axis_0 = const()[name = string("op_1015_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_1 = split(axis = var_1015_axis_0, split_sizes = tile_3, x = var_1012_cast_fp16)[name = string("op_1015_cast_fp16")];
            tensor<int32, [2]> var_1018_split_sizes_0 = const()[name = string("op_1018_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1018_axis_0 = const()[name = string("op_1018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1018_0, tensor<fp16, [1, 8, 128, 8]> var_1018_1 = split(axis = var_1018_axis_0, split_sizes = var_1018_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1018")];
            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_1005_cast_fp16_0, y = var_1018_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1021_to_fp16 = const()[name = string("op_1021_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_1021_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_1025 = const()[name = string("op_1025"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_23_cast_fp16 = softmax(axis = var_1025, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1031_transpose_x_1 = const()[name = string("op_1031_transpose_x_1"), val = bool(true)];
            bool var_1031_transpose_y_1 = const()[name = string("op_1031_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1031_cast_fp16 = matmul(transpose_x = var_1031_transpose_x_1, transpose_y = var_1031_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1015_cast_fp16_0)[name = string("op_1031_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_1005_cast_fp16_1, y = var_1018_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1033_to_fp16 = const()[name = string("op_1033_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_1033_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_1037 = const()[name = string("op_1037"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_31_cast_fp16 = softmax(axis = var_1037, 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_1015_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1045 = const()[name = string("op_1045"), 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_1045, interleave = attn_output_11_interleave_0, values = (var_1031_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1049_perm_0 = const()[name = string("op_1049_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_1049_cast_fp16 = transpose(perm = var_1049_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1049_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553154624)))];
            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_to_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_1082_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1082_cast_fp16")];
            int32 var_1080 = const()[name = string("op_1080"), 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_1080, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1082_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(561543296)))];
            fp16 var_1092_to_fp16 = const()[name = string("op_1092_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_1092_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1103_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1103_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561551552)))];
            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_to_fp16, x = var_1103_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1120_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [2]> var_1126_strides_0 = const()[name = string("op_1126_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1126_pad_type_0 = const()[name = string("op_1126_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1126_pad_0 = const()[name = string("op_1126_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1126_dilations_0 = const()[name = string("op_1126_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1126_groups_0 = const()[name = string("op_1126_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1126_cast_fp16 = conv(dilations = var_1126_dilations_0, groups = var_1126_groups_0, pad = var_1126_pad_0, pad_type = var_1126_pad_type_0, strides = var_1126_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1103_cast_fp16_0)[name = string("op_1126_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_19_cast_fp16 = mul(x = var_1120_cast_fp16, y = var_1126_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586717440)))];
            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_to_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_1144_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1144_cast_fp16")];
            int32 var_1142 = const()[name = string("op_1142"), 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_1142, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1144_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(611883328)))];
            fp16 var_1154_to_fp16 = const()[name = string("op_1154_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_1154_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1165_split_sizes_0 = const()[name = string("op_1165_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1165_axis_0 = const()[name = string("op_1165_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1165_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1165_cast_fp16_1 = split(axis = var_1165_axis_0, split_sizes = var_1165_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1165_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611891584)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620280256)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            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_cast_fp16, x = var_1165_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_1222_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1222_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_1229_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1229_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1233_cast_fp16 = mul(x = x_21_cast_fp16, y = var_452_cast_fp16)[name = string("op_1233_cast_fp16")];
            tensor<int32, [2]> var_1234_split_sizes_0 = const()[name = string("op_1234_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1234_axis_0 = const()[name = string("op_1234_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1234_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1234_cast_fp16_1 = split(axis = var_1234_axis_0, split_sizes = var_1234_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1234_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_1236_cast_fp16 = mul(x = var_1234_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1236_cast_fp16")];
            int32 var_1238 = const()[name = string("op_1238"), val = int32(-2)];
            bool var_1239_interleave_0 = const()[name = string("op_1239_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1239_cast_fp16 = concat(axis = var_1238, interleave = var_1239_interleave_0, values = (var_1236_cast_fp16, var_1234_cast_fp16_0))[name = string("op_1239_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1240_cast_fp16 = mul(x = var_1239_cast_fp16, y = var_459_cast_fp16)[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_15_cast_fp16 = add(x = var_1233_cast_fp16, y = var_1240_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1246_cast_fp16 = mul(x = var_1222_cast_fp16, y = var_452_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<int32, [2]> var_1247_split_sizes_0 = const()[name = string("op_1247_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1247_axis_0 = const()[name = string("op_1247_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1247_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1247_cast_fp16_1 = split(axis = var_1247_axis_0, split_sizes = var_1247_split_sizes_0, x = var_1222_cast_fp16)[name = string("op_1247_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_1249_cast_fp16 = mul(x = var_1247_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1249_cast_fp16")];
            int32 var_1251 = const()[name = string("op_1251"), val = int32(-2)];
            bool var_1252_interleave_0 = const()[name = string("op_1252_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1252_cast_fp16 = concat(axis = var_1251, interleave = var_1252_interleave_0, values = (var_1249_cast_fp16, var_1247_cast_fp16_0))[name = string("op_1252_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1253_cast_fp16 = mul(x = var_1252_cast_fp16, y = var_459_cast_fp16)[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_25_cast_fp16 = add(x = var_1246_cast_fp16, y = var_1253_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_125")];
            tensor<fp16, [14, 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_58)[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_60_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_60 = read_state(input = key_cache)[name = string("coreml_update_state_60")];
            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_1229_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [14, 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_59)[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_61_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_61 = read_state(input = value_cache)[name = string("coreml_update_state_61")];
            tensor<int32, [4]> var_1323_begin_0 = const()[name = string("op_1323_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1323_end_0 = const()[name = string("op_1323_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1323_end_mask_0 = const()[name = string("op_1323_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1323_cast_fp16 = slice_by_index(begin = var_1323_begin_0, end = var_1323_end_0, end_mask = var_1323_end_mask_0, x = coreml_update_state_60)[name = string("op_1323_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1326_axis_0 = const()[name = string("op_1326_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_1 = split(axis = var_1326_axis_0, split_sizes = tile_4, x = var_1323_cast_fp16)[name = string("op_1326_cast_fp16")];
            tensor<int32, [4]> var_1333_begin_0 = const()[name = string("op_1333_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1333_end_0 = const()[name = string("op_1333_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1333_end_mask_0 = const()[name = string("op_1333_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1333_cast_fp16 = slice_by_index(begin = var_1333_begin_0, end = var_1333_end_0, end_mask = var_1333_end_mask_0, x = coreml_update_state_61)[name = string("op_1333_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1336_axis_0 = const()[name = string("op_1336_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_1 = split(axis = var_1336_axis_0, split_sizes = tile_5, x = var_1333_cast_fp16)[name = string("op_1336_cast_fp16")];
            tensor<int32, [2]> var_1339_split_sizes_0 = const()[name = string("op_1339_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1339_axis_0 = const()[name = string("op_1339_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1339_0, tensor<fp16, [1, 8, 128, 8]> var_1339_1 = split(axis = var_1339_axis_0, split_sizes = var_1339_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1339")];
            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_1326_cast_fp16_0, y = var_1339_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1342_to_fp16 = const()[name = string("op_1342_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_1342_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_1346 = const()[name = string("op_1346"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_39_cast_fp16 = softmax(axis = var_1346, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1352_transpose_x_1 = const()[name = string("op_1352_transpose_x_1"), val = bool(true)];
            bool var_1352_transpose_y_1 = const()[name = string("op_1352_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1352_cast_fp16 = matmul(transpose_x = var_1352_transpose_x_1, transpose_y = var_1352_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1336_cast_fp16_0)[name = string("op_1352_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_1326_cast_fp16_1, y = var_1339_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1354_to_fp16 = const()[name = string("op_1354_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_1354_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_1358 = const()[name = string("op_1358"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_47_cast_fp16 = softmax(axis = var_1358, 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_1336_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1366 = const()[name = string("op_1366"), 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_1366, interleave = attn_output_19_interleave_0, values = (var_1352_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1370_perm_0 = const()[name = string("op_1370_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_1370_cast_fp16 = transpose(perm = var_1370_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1370_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621328896)))];
            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_to_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_1403_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1403_cast_fp16")];
            int32 var_1401 = const()[name = string("op_1401"), 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_1401, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1403_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(629717568)))];
            fp16 var_1413_to_fp16 = const()[name = string("op_1413_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_1413_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1424_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1424_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629725824)))];
            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_to_fp16, x = var_1424_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1441_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1441_cast_fp16")];
            tensor<int32, [2]> var_1447_strides_0 = const()[name = string("op_1447_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1447_pad_type_0 = const()[name = string("op_1447_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1447_pad_0 = const()[name = string("op_1447_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1447_dilations_0 = const()[name = string("op_1447_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1447_groups_0 = const()[name = string("op_1447_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1447_cast_fp16 = conv(dilations = var_1447_dilations_0, groups = var_1447_groups_0, pad = var_1447_pad_0, pad_type = var_1447_pad_type_0, strides = var_1447_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1424_cast_fp16_0)[name = string("op_1447_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_29_cast_fp16 = mul(x = var_1441_cast_fp16, y = var_1447_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654891712)))];
            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_to_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_1465_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1465_cast_fp16")];
            int32 var_1463 = const()[name = string("op_1463"), 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_1463, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1465_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(680057600)))];
            fp16 var_1475_to_fp16 = const()[name = string("op_1475_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_1475_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1486_split_sizes_0 = const()[name = string("op_1486_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1486_axis_0 = const()[name = string("op_1486_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1486_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1486_cast_fp16_1 = split(axis = var_1486_axis_0, split_sizes = var_1486_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1486_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680065856)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688454528)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            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_cast_fp16, x = var_1486_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_1543_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1543_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_1550_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1554_cast_fp16 = mul(x = x_31_cast_fp16, y = var_452_cast_fp16)[name = string("op_1554_cast_fp16")];
            tensor<int32, [2]> var_1555_split_sizes_0 = const()[name = string("op_1555_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1555_axis_0 = const()[name = string("op_1555_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1555_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1555_cast_fp16_1 = split(axis = var_1555_axis_0, split_sizes = var_1555_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1555_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_1557_cast_fp16 = mul(x = var_1555_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1557_cast_fp16")];
            int32 var_1559 = const()[name = string("op_1559"), val = int32(-2)];
            bool var_1560_interleave_0 = const()[name = string("op_1560_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1560_cast_fp16 = concat(axis = var_1559, interleave = var_1560_interleave_0, values = (var_1557_cast_fp16, var_1555_cast_fp16_0))[name = string("op_1560_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1561_cast_fp16 = mul(x = var_1560_cast_fp16, y = var_459_cast_fp16)[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_21_cast_fp16 = add(x = var_1554_cast_fp16, y = var_1561_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1567_cast_fp16 = mul(x = var_1543_cast_fp16, y = var_452_cast_fp16)[name = string("op_1567_cast_fp16")];
            tensor<int32, [2]> var_1568_split_sizes_0 = const()[name = string("op_1568_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1568_axis_0 = const()[name = string("op_1568_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1568_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1568_cast_fp16_1 = split(axis = var_1568_axis_0, split_sizes = var_1568_split_sizes_0, x = var_1543_cast_fp16)[name = string("op_1568_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_1570_cast_fp16 = mul(x = var_1568_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1570_cast_fp16")];
            int32 var_1572 = const()[name = string("op_1572"), val = int32(-2)];
            bool var_1573_interleave_0 = const()[name = string("op_1573_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1573_cast_fp16 = concat(axis = var_1572, interleave = var_1573_interleave_0, values = (var_1570_cast_fp16, var_1568_cast_fp16_0))[name = string("op_1573_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1574_cast_fp16 = mul(x = var_1573_cast_fp16, y = var_459_cast_fp16)[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_35_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1574_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_122")];
            tensor<fp16, [14, 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_60)[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_62_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_62 = read_state(input = key_cache)[name = string("coreml_update_state_62")];
            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_1550_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [14, 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_61)[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_63_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_63 = read_state(input = value_cache)[name = string("coreml_update_state_63")];
            tensor<int32, [4]> var_1644_begin_0 = const()[name = string("op_1644_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1644_end_0 = const()[name = string("op_1644_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1644_end_mask_0 = const()[name = string("op_1644_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1644_cast_fp16 = slice_by_index(begin = var_1644_begin_0, end = var_1644_end_0, end_mask = var_1644_end_mask_0, x = coreml_update_state_62)[name = string("op_1644_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1647_axis_0 = const()[name = string("op_1647_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_1 = split(axis = var_1647_axis_0, split_sizes = tile_6, x = var_1644_cast_fp16)[name = string("op_1647_cast_fp16")];
            tensor<int32, [4]> var_1654_begin_0 = const()[name = string("op_1654_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1654_end_0 = const()[name = string("op_1654_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1654_end_mask_0 = const()[name = string("op_1654_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1654_cast_fp16 = slice_by_index(begin = var_1654_begin_0, end = var_1654_end_0, end_mask = var_1654_end_mask_0, x = coreml_update_state_63)[name = string("op_1654_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1657_axis_0 = const()[name = string("op_1657_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_1 = split(axis = var_1657_axis_0, split_sizes = tile_7, x = var_1654_cast_fp16)[name = string("op_1657_cast_fp16")];
            tensor<int32, [2]> var_1660_split_sizes_0 = const()[name = string("op_1660_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1660_axis_0 = const()[name = string("op_1660_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1660_0, tensor<fp16, [1, 8, 128, 8]> var_1660_1 = split(axis = var_1660_axis_0, split_sizes = var_1660_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1660")];
            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_1647_cast_fp16_0, y = var_1660_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1663_to_fp16 = const()[name = string("op_1663_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_1663_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_1667 = const()[name = string("op_1667"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_55_cast_fp16 = softmax(axis = var_1667, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1673_transpose_x_1 = const()[name = string("op_1673_transpose_x_1"), val = bool(true)];
            bool var_1673_transpose_y_1 = const()[name = string("op_1673_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1673_cast_fp16 = matmul(transpose_x = var_1673_transpose_x_1, transpose_y = var_1673_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1657_cast_fp16_0)[name = string("op_1673_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_1647_cast_fp16_1, y = var_1660_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1675_to_fp16 = const()[name = string("op_1675_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_1675_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_1679 = const()[name = string("op_1679"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_63_cast_fp16 = softmax(axis = var_1679, 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_1657_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1687 = const()[name = string("op_1687"), 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_1687, interleave = attn_output_27_interleave_0, values = (var_1673_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1691_perm_0 = const()[name = string("op_1691_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_1691_cast_fp16 = transpose(perm = var_1691_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1691_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            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_cast_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_1724_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1724_cast_fp16")];
            int32 var_1722 = const()[name = string("op_1722"), 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_1722, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1724_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(689503168)))];
            fp16 var_1734_to_fp16 = const()[name = string("op_1734_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_1734_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1745_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1745_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689511424)))];
            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_to_fp16, x = var_1745_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1762_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1762_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(714677312)))];
            tensor<int32, [2]> var_1768_strides_0 = const()[name = string("op_1768_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1768_pad_type_0 = const()[name = string("op_1768_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1768_pad_0 = const()[name = string("op_1768_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1768_dilations_0 = const()[name = string("op_1768_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1768_groups_0 = const()[name = string("op_1768_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1768_cast_fp16 = conv(dilations = var_1768_dilations_0, groups = var_1768_groups_0, pad = var_1768_pad_0, pad_type = var_1768_pad_type_0, strides = var_1768_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1745_cast_fp16_0)[name = string("op_1768_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_39_cast_fp16 = mul(x = var_1762_cast_fp16, y = var_1768_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_1786_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1786_cast_fp16")];
            int32 var_1784 = const()[name = string("op_1784"), 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_1784, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1786_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(739843200)))];
            fp16 var_1796_to_fp16 = const()[name = string("op_1796_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_1796_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1807_split_sizes_0 = const()[name = string("op_1807_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1807_axis_0 = const()[name = string("op_1807_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1807_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1807_cast_fp16_1 = split(axis = var_1807_axis_0, split_sizes = var_1807_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739851456)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(748240128)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            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_cast_fp16, x = var_1807_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_1864_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1864_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_1871_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1871_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1875_cast_fp16 = mul(x = x_41_cast_fp16, y = var_452_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [2]> var_1876_split_sizes_0 = const()[name = string("op_1876_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1876_axis_0 = const()[name = string("op_1876_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1876_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1876_cast_fp16_1 = split(axis = var_1876_axis_0, split_sizes = var_1876_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1876_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_1878_cast_fp16 = mul(x = var_1876_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1878_cast_fp16")];
            int32 var_1880 = const()[name = string("op_1880"), val = int32(-2)];
            bool var_1881_interleave_0 = const()[name = string("op_1881_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1881_cast_fp16 = concat(axis = var_1880, interleave = var_1881_interleave_0, values = (var_1878_cast_fp16, var_1876_cast_fp16_0))[name = string("op_1881_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1882_cast_fp16 = mul(x = var_1881_cast_fp16, y = var_459_cast_fp16)[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_27_cast_fp16 = add(x = var_1875_cast_fp16, y = var_1882_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1888_cast_fp16 = mul(x = var_1864_cast_fp16, y = var_452_cast_fp16)[name = string("op_1888_cast_fp16")];
            tensor<int32, [2]> var_1889_split_sizes_0 = const()[name = string("op_1889_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1889_axis_0 = const()[name = string("op_1889_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1889_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1889_cast_fp16_1 = split(axis = var_1889_axis_0, split_sizes = var_1889_split_sizes_0, x = var_1864_cast_fp16)[name = string("op_1889_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_1891_cast_fp16 = mul(x = var_1889_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1891_cast_fp16")];
            int32 var_1893 = const()[name = string("op_1893"), val = int32(-2)];
            bool var_1894_interleave_0 = const()[name = string("op_1894_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1894_cast_fp16 = concat(axis = var_1893, interleave = var_1894_interleave_0, values = (var_1891_cast_fp16, var_1889_cast_fp16_0))[name = string("op_1894_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1895_cast_fp16 = mul(x = var_1894_cast_fp16, y = var_459_cast_fp16)[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_45_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1895_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_119")];
            tensor<fp16, [14, 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_62)[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_64_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_64 = read_state(input = key_cache)[name = string("coreml_update_state_64")];
            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_1871_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [14, 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_63)[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_65_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_65 = read_state(input = value_cache)[name = string("coreml_update_state_65")];
            tensor<int32, [4]> var_1965_begin_0 = const()[name = string("op_1965_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1965_end_0 = const()[name = string("op_1965_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1965_end_mask_0 = const()[name = string("op_1965_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1965_cast_fp16 = slice_by_index(begin = var_1965_begin_0, end = var_1965_end_0, end_mask = var_1965_end_mask_0, x = coreml_update_state_64)[name = string("op_1965_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1968_axis_0 = const()[name = string("op_1968_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_1 = split(axis = var_1968_axis_0, split_sizes = tile_8, x = var_1965_cast_fp16)[name = string("op_1968_cast_fp16")];
            tensor<int32, [4]> var_1975_begin_0 = const()[name = string("op_1975_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1975_end_0 = const()[name = string("op_1975_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1975_end_mask_0 = const()[name = string("op_1975_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1975_cast_fp16 = slice_by_index(begin = var_1975_begin_0, end = var_1975_end_0, end_mask = var_1975_end_mask_0, x = coreml_update_state_65)[name = string("op_1975_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1978_axis_0 = const()[name = string("op_1978_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_1 = split(axis = var_1978_axis_0, split_sizes = tile_9, x = var_1975_cast_fp16)[name = string("op_1978_cast_fp16")];
            tensor<int32, [2]> var_1981_split_sizes_0 = const()[name = string("op_1981_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1981_axis_0 = const()[name = string("op_1981_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1981_0, tensor<fp16, [1, 8, 128, 8]> var_1981_1 = split(axis = var_1981_axis_0, split_sizes = var_1981_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1981")];
            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_1968_cast_fp16_0, y = var_1981_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1984_to_fp16 = const()[name = string("op_1984_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_1984_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_1988 = const()[name = string("op_1988"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_71_cast_fp16 = softmax(axis = var_1988, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1994_transpose_x_1 = const()[name = string("op_1994_transpose_x_1"), val = bool(true)];
            bool var_1994_transpose_y_1 = const()[name = string("op_1994_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1994_cast_fp16 = matmul(transpose_x = var_1994_transpose_x_1, transpose_y = var_1994_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1978_cast_fp16_0)[name = string("op_1994_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_1968_cast_fp16_1, y = var_1981_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1996_to_fp16 = const()[name = string("op_1996_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_1996_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_2000 = const()[name = string("op_2000"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_79_cast_fp16 = softmax(axis = var_2000, 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_1978_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2008 = const()[name = string("op_2008"), 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_2008, interleave = attn_output_35_interleave_0, values = (var_1994_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2012_perm_0 = const()[name = string("op_2012_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_2012_cast_fp16 = transpose(perm = var_2012_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2012_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_2045_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2045_cast_fp16")];
            int32 var_2043 = const()[name = string("op_2043"), 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_2043, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2045_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(749288768)))];
            fp16 var_2055_to_fp16 = const()[name = string("op_2055_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_2055_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2066_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2066_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_2066_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2083_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2083_cast_fp16")];
            tensor<int32, [2]> var_2089_strides_0 = const()[name = string("op_2089_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2089_pad_type_0 = const()[name = string("op_2089_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2089_pad_0 = const()[name = string("op_2089_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2089_dilations_0 = const()[name = string("op_2089_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2089_groups_0 = const()[name = string("op_2089_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2089_cast_fp16 = conv(dilations = var_2089_dilations_0, groups = var_2089_groups_0, pad = var_2089_pad_0, pad_type = var_2089_pad_type_0, strides = var_2089_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2066_cast_fp16_0)[name = string("op_2089_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_49_cast_fp16 = mul(x = var_2083_cast_fp16, y = var_2089_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_2107_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2107_cast_fp16")];
            int32 var_2105 = const()[name = string("op_2105"), 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_2105, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2107_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(749297024)))];
            fp16 var_2117_to_fp16 = const()[name = string("op_2117_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_2117_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2128_split_sizes_0 = const()[name = string("op_2128_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2128_axis_0 = const()[name = string("op_2128_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2128_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2128_cast_fp16_1 = split(axis = var_2128_axis_0, split_sizes = var_2128_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749305280)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(757693952)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            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_cast_fp16, x = var_2128_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_2185_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2185_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_2192_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2196_cast_fp16 = mul(x = x_51_cast_fp16, y = var_452_cast_fp16)[name = string("op_2196_cast_fp16")];
            tensor<int32, [2]> var_2197_split_sizes_0 = const()[name = string("op_2197_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2197_axis_0 = const()[name = string("op_2197_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2197_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2197_cast_fp16_1 = split(axis = var_2197_axis_0, split_sizes = var_2197_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2197_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_2199_cast_fp16 = mul(x = var_2197_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2199_cast_fp16")];
            int32 var_2201 = const()[name = string("op_2201"), val = int32(-2)];
            bool var_2202_interleave_0 = const()[name = string("op_2202_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2202_cast_fp16 = concat(axis = var_2201, interleave = var_2202_interleave_0, values = (var_2199_cast_fp16, var_2197_cast_fp16_0))[name = string("op_2202_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2203_cast_fp16 = mul(x = var_2202_cast_fp16, y = var_459_cast_fp16)[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_33_cast_fp16 = add(x = var_2196_cast_fp16, y = var_2203_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2209_cast_fp16 = mul(x = var_2185_cast_fp16, y = var_452_cast_fp16)[name = string("op_2209_cast_fp16")];
            tensor<int32, [2]> var_2210_split_sizes_0 = const()[name = string("op_2210_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2210_axis_0 = const()[name = string("op_2210_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2210_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2210_cast_fp16_1 = split(axis = var_2210_axis_0, split_sizes = var_2210_split_sizes_0, x = var_2185_cast_fp16)[name = string("op_2210_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_2212_cast_fp16 = mul(x = var_2210_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2212_cast_fp16")];
            int32 var_2214 = const()[name = string("op_2214"), val = int32(-2)];
            bool var_2215_interleave_0 = const()[name = string("op_2215_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2215_cast_fp16 = concat(axis = var_2214, interleave = var_2215_interleave_0, values = (var_2212_cast_fp16, var_2210_cast_fp16_0))[name = string("op_2215_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2216_cast_fp16 = mul(x = var_2215_cast_fp16, y = var_459_cast_fp16)[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_55_cast_fp16 = add(x = var_2209_cast_fp16, y = var_2216_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_116")];
            tensor<fp16, [14, 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_64)[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_66_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_66 = read_state(input = key_cache)[name = string("coreml_update_state_66")];
            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_2192_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [14, 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_65)[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_67_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_67 = read_state(input = value_cache)[name = string("coreml_update_state_67")];
            tensor<int32, [4]> var_2286_begin_0 = const()[name = string("op_2286_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2286_end_0 = const()[name = string("op_2286_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2286_end_mask_0 = const()[name = string("op_2286_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2286_cast_fp16 = slice_by_index(begin = var_2286_begin_0, end = var_2286_end_0, end_mask = var_2286_end_mask_0, x = coreml_update_state_66)[name = string("op_2286_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2289_axis_0 = const()[name = string("op_2289_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_1 = split(axis = var_2289_axis_0, split_sizes = tile_10, x = var_2286_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> var_2296_begin_0 = const()[name = string("op_2296_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2296_end_0 = const()[name = string("op_2296_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2296_end_mask_0 = const()[name = string("op_2296_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2296_cast_fp16 = slice_by_index(begin = var_2296_begin_0, end = var_2296_end_0, end_mask = var_2296_end_mask_0, x = coreml_update_state_67)[name = string("op_2296_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2299_axis_0 = const()[name = string("op_2299_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_1 = split(axis = var_2299_axis_0, split_sizes = tile_11, x = var_2296_cast_fp16)[name = string("op_2299_cast_fp16")];
            tensor<int32, [2]> var_2302_split_sizes_0 = const()[name = string("op_2302_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2302_axis_0 = const()[name = string("op_2302_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2302_0, tensor<fp16, [1, 8, 128, 8]> var_2302_1 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2302")];
            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_2289_cast_fp16_0, y = var_2302_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2305_to_fp16 = const()[name = string("op_2305_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_2305_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_2309 = const()[name = string("op_2309"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_87_cast_fp16 = softmax(axis = var_2309, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2315_transpose_x_1 = const()[name = string("op_2315_transpose_x_1"), val = bool(true)];
            bool var_2315_transpose_y_1 = const()[name = string("op_2315_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2315_cast_fp16 = matmul(transpose_x = var_2315_transpose_x_1, transpose_y = var_2315_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2299_cast_fp16_0)[name = string("op_2315_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_2289_cast_fp16_1, y = var_2302_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2317_to_fp16 = const()[name = string("op_2317_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_2317_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_2321 = const()[name = string("op_2321"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_95_cast_fp16 = softmax(axis = var_2321, 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_2299_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2329 = const()[name = string("op_2329"), 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_2329, interleave = attn_output_43_interleave_0, values = (var_2315_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2333_perm_0 = const()[name = string("op_2333_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_2333_cast_fp16 = transpose(perm = var_2333_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2333_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_2366_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2366_cast_fp16")];
            int32 var_2364 = const()[name = string("op_2364"), 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_2364, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2366_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(758742592)))];
            fp16 var_2376_to_fp16 = const()[name = string("op_2376_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_2376_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2387_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2387_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(758750848)))];
            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_to_fp16, x = var_2387_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2404_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [2]> var_2410_strides_0 = const()[name = string("op_2410_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2410_pad_type_0 = const()[name = string("op_2410_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2410_pad_0 = const()[name = string("op_2410_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2410_dilations_0 = const()[name = string("op_2410_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2410_groups_0 = const()[name = string("op_2410_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2410_cast_fp16 = conv(dilations = var_2410_dilations_0, groups = var_2410_groups_0, pad = var_2410_pad_0, pad_type = var_2410_pad_type_0, strides = var_2410_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2387_cast_fp16_0)[name = string("op_2410_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_59_cast_fp16 = mul(x = var_2404_cast_fp16, y = var_2410_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_2428_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2428_cast_fp16")];
            int32 var_2426 = const()[name = string("op_2426"), 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_2426, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2428_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(783916736)))];
            fp16 var_2438_to_fp16 = const()[name = string("op_2438_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_2438_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2449_split_sizes_0 = const()[name = string("op_2449_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2449_axis_0 = const()[name = string("op_2449_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2449_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2449_cast_fp16_1 = split(axis = var_2449_axis_0, split_sizes = var_2449_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2449_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(783924992)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792313664)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            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_cast_fp16, x = var_2449_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_2506_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2506_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_2513_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2517_cast_fp16 = mul(x = x_61_cast_fp16, y = var_452_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<int32, [2]> var_2518_split_sizes_0 = const()[name = string("op_2518_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2518_axis_0 = const()[name = string("op_2518_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2518_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2518_cast_fp16_1 = split(axis = var_2518_axis_0, split_sizes = var_2518_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2518_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_2520_cast_fp16 = mul(x = var_2518_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2520_cast_fp16")];
            int32 var_2522 = const()[name = string("op_2522"), val = int32(-2)];
            bool var_2523_interleave_0 = const()[name = string("op_2523_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2523_cast_fp16 = concat(axis = var_2522, interleave = var_2523_interleave_0, values = (var_2520_cast_fp16, var_2518_cast_fp16_0))[name = string("op_2523_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2524_cast_fp16 = mul(x = var_2523_cast_fp16, y = var_459_cast_fp16)[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_39_cast_fp16 = add(x = var_2517_cast_fp16, y = var_2524_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2530_cast_fp16 = mul(x = var_2506_cast_fp16, y = var_452_cast_fp16)[name = string("op_2530_cast_fp16")];
            tensor<int32, [2]> var_2531_split_sizes_0 = const()[name = string("op_2531_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2531_axis_0 = const()[name = string("op_2531_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2531_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2531_cast_fp16_1 = split(axis = var_2531_axis_0, split_sizes = var_2531_split_sizes_0, x = var_2506_cast_fp16)[name = string("op_2531_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_2533_cast_fp16 = mul(x = var_2531_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2533_cast_fp16")];
            int32 var_2535 = const()[name = string("op_2535"), val = int32(-2)];
            bool var_2536_interleave_0 = const()[name = string("op_2536_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2536_cast_fp16 = concat(axis = var_2535, interleave = var_2536_interleave_0, values = (var_2533_cast_fp16, var_2531_cast_fp16_0))[name = string("op_2536_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2537_cast_fp16 = mul(x = var_2536_cast_fp16, y = var_459_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_65_cast_fp16 = add(x = var_2530_cast_fp16, y = var_2537_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_113")];
            tensor<fp16, [14, 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_66)[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_68_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_68 = read_state(input = key_cache)[name = string("coreml_update_state_68")];
            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_2513_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [14, 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_67)[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_69_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_69 = read_state(input = value_cache)[name = string("coreml_update_state_69")];
            tensor<int32, [4]> var_2607_begin_0 = const()[name = string("op_2607_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2607_end_0 = const()[name = string("op_2607_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2607_end_mask_0 = const()[name = string("op_2607_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2607_cast_fp16 = slice_by_index(begin = var_2607_begin_0, end = var_2607_end_0, end_mask = var_2607_end_mask_0, x = coreml_update_state_68)[name = string("op_2607_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2610_axis_0 = const()[name = string("op_2610_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_1 = split(axis = var_2610_axis_0, split_sizes = tile_12, x = var_2607_cast_fp16)[name = string("op_2610_cast_fp16")];
            tensor<int32, [4]> var_2617_begin_0 = const()[name = string("op_2617_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2617_end_0 = const()[name = string("op_2617_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2617_end_mask_0 = const()[name = string("op_2617_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2617_cast_fp16 = slice_by_index(begin = var_2617_begin_0, end = var_2617_end_0, end_mask = var_2617_end_mask_0, x = coreml_update_state_69)[name = string("op_2617_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2620_axis_0 = const()[name = string("op_2620_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_1 = split(axis = var_2620_axis_0, split_sizes = tile_13, x = var_2617_cast_fp16)[name = string("op_2620_cast_fp16")];
            tensor<int32, [2]> var_2623_split_sizes_0 = const()[name = string("op_2623_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2623_axis_0 = const()[name = string("op_2623_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2623_0, tensor<fp16, [1, 8, 128, 8]> var_2623_1 = split(axis = var_2623_axis_0, split_sizes = var_2623_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2623")];
            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_2610_cast_fp16_0, y = var_2623_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2626_to_fp16 = const()[name = string("op_2626_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_2626_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_2630 = const()[name = string("op_2630"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_103_cast_fp16 = softmax(axis = var_2630, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2636_transpose_x_1 = const()[name = string("op_2636_transpose_x_1"), val = bool(true)];
            bool var_2636_transpose_y_1 = const()[name = string("op_2636_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2636_cast_fp16 = matmul(transpose_x = var_2636_transpose_x_1, transpose_y = var_2636_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2620_cast_fp16_0)[name = string("op_2636_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_2610_cast_fp16_1, y = var_2623_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2638_to_fp16 = const()[name = string("op_2638_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_2638_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_2642 = const()[name = string("op_2642"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_111_cast_fp16 = softmax(axis = var_2642, 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_2620_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2650 = const()[name = string("op_2650"), 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_2650, interleave = attn_output_51_interleave_0, values = (var_2636_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2654_perm_0 = const()[name = string("op_2654_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_2654_cast_fp16 = transpose(perm = var_2654_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2654_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_2687_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2687_cast_fp16")];
            int32 var_2685 = const()[name = string("op_2685"), 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_2685, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2687_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(793362304)))];
            fp16 var_2697_to_fp16 = const()[name = string("op_2697_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_2697_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2708_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2708_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_2708_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2725_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2725_cast_fp16")];
            tensor<int32, [2]> var_2731_strides_0 = const()[name = string("op_2731_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2731_pad_type_0 = const()[name = string("op_2731_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2731_pad_0 = const()[name = string("op_2731_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2731_dilations_0 = const()[name = string("op_2731_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2731_groups_0 = const()[name = string("op_2731_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2731_cast_fp16 = conv(dilations = var_2731_dilations_0, groups = var_2731_groups_0, pad = var_2731_pad_0, pad_type = var_2731_pad_type_0, strides = var_2731_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2708_cast_fp16_0)[name = string("op_2731_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_69_cast_fp16 = mul(x = var_2725_cast_fp16, y = var_2731_cast_fp16)[name = string("x_69_cast_fp16")];
            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_cast_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_2749_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2749_cast_fp16")];
            int32 var_2747 = const()[name = string("op_2747"), 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_2747, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2749_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(793370560)))];
            fp16 var_2759_to_fp16 = const()[name = string("op_2759_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_2759_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2770_split_sizes_0 = const()[name = string("op_2770_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2770_axis_0 = const()[name = string("op_2770_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2770_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2770_cast_fp16_1 = split(axis = var_2770_axis_0, split_sizes = var_2770_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2770_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(793378816)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(801767488)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            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_cast_fp16, x = var_2770_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_2827_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2827_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_2834_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2834_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2838_cast_fp16 = mul(x = x_71_cast_fp16, y = var_452_cast_fp16)[name = string("op_2838_cast_fp16")];
            tensor<int32, [2]> var_2839_split_sizes_0 = const()[name = string("op_2839_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2839_axis_0 = const()[name = string("op_2839_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2839_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2839_cast_fp16_1 = split(axis = var_2839_axis_0, split_sizes = var_2839_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2839_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_2841_cast_fp16 = mul(x = var_2839_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2841_cast_fp16")];
            int32 var_2843 = const()[name = string("op_2843"), val = int32(-2)];
            bool var_2844_interleave_0 = const()[name = string("op_2844_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2844_cast_fp16 = concat(axis = var_2843, interleave = var_2844_interleave_0, values = (var_2841_cast_fp16, var_2839_cast_fp16_0))[name = string("op_2844_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2845_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_459_cast_fp16)[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_45_cast_fp16 = add(x = var_2838_cast_fp16, y = var_2845_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2851_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_452_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [2]> var_2852_split_sizes_0 = const()[name = string("op_2852_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2852_axis_0 = const()[name = string("op_2852_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2852_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2852_cast_fp16_1 = split(axis = var_2852_axis_0, split_sizes = var_2852_split_sizes_0, x = var_2827_cast_fp16)[name = string("op_2852_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_2854_cast_fp16 = mul(x = var_2852_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2854_cast_fp16")];
            int32 var_2856 = const()[name = string("op_2856"), val = int32(-2)];
            bool var_2857_interleave_0 = const()[name = string("op_2857_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2857_cast_fp16 = concat(axis = var_2856, interleave = var_2857_interleave_0, values = (var_2854_cast_fp16, var_2852_cast_fp16_0))[name = string("op_2857_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2858_cast_fp16 = mul(x = var_2857_cast_fp16, y = var_459_cast_fp16)[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_75_cast_fp16 = add(x = var_2851_cast_fp16, y = var_2858_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_110")];
            tensor<fp16, [14, 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_68)[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_70_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_70 = read_state(input = key_cache)[name = string("coreml_update_state_70")];
            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_2834_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [14, 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_69)[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_71_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_71 = read_state(input = value_cache)[name = string("coreml_update_state_71")];
            tensor<int32, [4]> var_2928_begin_0 = const()[name = string("op_2928_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2928_end_0 = const()[name = string("op_2928_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2928_end_mask_0 = const()[name = string("op_2928_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2928_cast_fp16 = slice_by_index(begin = var_2928_begin_0, end = var_2928_end_0, end_mask = var_2928_end_mask_0, x = coreml_update_state_70)[name = string("op_2928_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2931_axis_0 = const()[name = string("op_2931_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_1 = split(axis = var_2931_axis_0, split_sizes = tile_14, x = var_2928_cast_fp16)[name = string("op_2931_cast_fp16")];
            tensor<int32, [4]> var_2938_begin_0 = const()[name = string("op_2938_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2938_end_0 = const()[name = string("op_2938_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2938_end_mask_0 = const()[name = string("op_2938_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2938_cast_fp16 = slice_by_index(begin = var_2938_begin_0, end = var_2938_end_0, end_mask = var_2938_end_mask_0, x = coreml_update_state_71)[name = string("op_2938_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2941_axis_0 = const()[name = string("op_2941_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_1 = split(axis = var_2941_axis_0, split_sizes = tile_15, x = var_2938_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<int32, [2]> var_2944_split_sizes_0 = const()[name = string("op_2944_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2944_axis_0 = const()[name = string("op_2944_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2944_0, tensor<fp16, [1, 8, 128, 8]> var_2944_1 = split(axis = var_2944_axis_0, split_sizes = var_2944_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2944")];
            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_2931_cast_fp16_0, y = var_2944_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2947_to_fp16 = const()[name = string("op_2947_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_2947_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_2951 = const()[name = string("op_2951"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_119_cast_fp16 = softmax(axis = var_2951, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2957_transpose_x_1 = const()[name = string("op_2957_transpose_x_1"), val = bool(true)];
            bool var_2957_transpose_y_1 = const()[name = string("op_2957_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2957_cast_fp16 = matmul(transpose_x = var_2957_transpose_x_1, transpose_y = var_2957_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2941_cast_fp16_0)[name = string("op_2957_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_2931_cast_fp16_1, y = var_2944_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2959_to_fp16 = const()[name = string("op_2959_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_2959_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_2963 = const()[name = string("op_2963"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_127_cast_fp16 = softmax(axis = var_2963, 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_2941_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2971 = const()[name = string("op_2971"), 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_2971, interleave = attn_output_59_interleave_0, values = (var_2957_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2975_perm_0 = const()[name = string("op_2975_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_2975_cast_fp16 = transpose(perm = var_2975_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2975_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_3008_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3008_cast_fp16")];
            int32 var_3006 = const()[name = string("op_3006"), 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_3006, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3008_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(802816128)))];
            fp16 var_3018_to_fp16 = const()[name = string("op_3018_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_3018_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3029_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3029_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_3029_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_3046_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3046_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(802824384)))];
            tensor<int32, [2]> var_3052_strides_0 = const()[name = string("op_3052_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3052_pad_type_0 = const()[name = string("op_3052_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3052_pad_0 = const()[name = string("op_3052_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3052_dilations_0 = const()[name = string("op_3052_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3052_groups_0 = const()[name = string("op_3052_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_3052_cast_fp16 = conv(dilations = var_3052_dilations_0, groups = var_3052_groups_0, pad = var_3052_pad_0, pad_type = var_3052_pad_type_0, strides = var_3052_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_3029_cast_fp16_0)[name = string("op_3052_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_79_cast_fp16 = mul(x = var_3046_cast_fp16, y = var_3052_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(827990272)))];
            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_3070_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3070_cast_fp16")];
            int32 var_3068 = const()[name = string("op_3068"), 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_3068, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3070_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(853156160)))];
            fp16 var_3080_to_fp16 = const()[name = string("op_3080_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_3080_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3091_split_sizes_0 = const()[name = string("op_3091_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3091_axis_0 = const()[name = string("op_3091_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3091_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3091_cast_fp16_1 = split(axis = var_3091_axis_0, split_sizes = var_3091_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3091_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(853164416)))];
            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_3091_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(861553088)))];
            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_3091_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            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_cast_fp16, x = var_3091_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_3148_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3148_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_3155_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3155_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3159_cast_fp16 = mul(x = x_81_cast_fp16, y = var_452_cast_fp16)[name = string("op_3159_cast_fp16")];
            tensor<int32, [2]> var_3160_split_sizes_0 = const()[name = string("op_3160_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3160_axis_0 = const()[name = string("op_3160_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_3160_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_3160_cast_fp16_1 = split(axis = var_3160_axis_0, split_sizes = var_3160_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3160_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_3162_cast_fp16 = mul(x = var_3160_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3162_cast_fp16")];
            int32 var_3164 = const()[name = string("op_3164"), val = int32(-2)];
            bool var_3165_interleave_0 = const()[name = string("op_3165_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_3165_cast_fp16 = concat(axis = var_3164, interleave = var_3165_interleave_0, values = (var_3162_cast_fp16, var_3160_cast_fp16_0))[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3166_cast_fp16 = mul(x = var_3165_cast_fp16, y = var_459_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_51_cast_fp16 = add(x = var_3159_cast_fp16, y = var_3166_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3172_cast_fp16 = mul(x = var_3148_cast_fp16, y = var_452_cast_fp16)[name = string("op_3172_cast_fp16")];
            tensor<int32, [2]> var_3173_split_sizes_0 = const()[name = string("op_3173_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3173_axis_0 = const()[name = string("op_3173_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_3173_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_3173_cast_fp16_1 = split(axis = var_3173_axis_0, split_sizes = var_3173_split_sizes_0, x = var_3148_cast_fp16)[name = string("op_3173_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_3175_cast_fp16 = mul(x = var_3173_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3175_cast_fp16")];
            int32 var_3177 = const()[name = string("op_3177"), val = int32(-2)];
            bool var_3178_interleave_0 = const()[name = string("op_3178_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_3178_cast_fp16 = concat(axis = var_3177, interleave = var_3178_interleave_0, values = (var_3175_cast_fp16, var_3173_cast_fp16_0))[name = string("op_3178_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3179_cast_fp16 = mul(x = var_3178_cast_fp16, y = var_459_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_85_cast_fp16 = add(x = var_3172_cast_fp16, y = var_3179_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_107")];
            tensor<fp16, [14, 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_70)[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_72_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_72 = read_state(input = key_cache)[name = string("coreml_update_state_72")];
            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_3155_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [14, 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_71)[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_73_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_73 = read_state(input = value_cache)[name = string("coreml_update_state_73")];
            tensor<int32, [4]> var_3249_begin_0 = const()[name = string("op_3249_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3249_end_0 = const()[name = string("op_3249_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3249_end_mask_0 = const()[name = string("op_3249_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3249_cast_fp16 = slice_by_index(begin = var_3249_begin_0, end = var_3249_end_0, end_mask = var_3249_end_mask_0, x = coreml_update_state_72)[name = string("op_3249_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3252_axis_0 = const()[name = string("op_3252_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_1 = split(axis = var_3252_axis_0, split_sizes = tile_16, x = var_3249_cast_fp16)[name = string("op_3252_cast_fp16")];
            tensor<int32, [4]> var_3259_begin_0 = const()[name = string("op_3259_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3259_end_0 = const()[name = string("op_3259_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3259_end_mask_0 = const()[name = string("op_3259_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3259_cast_fp16 = slice_by_index(begin = var_3259_begin_0, end = var_3259_end_0, end_mask = var_3259_end_mask_0, x = coreml_update_state_73)[name = string("op_3259_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3262_axis_0 = const()[name = string("op_3262_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_1 = split(axis = var_3262_axis_0, split_sizes = tile_17, x = var_3259_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<int32, [2]> var_3265_split_sizes_0 = const()[name = string("op_3265_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3265_axis_0 = const()[name = string("op_3265_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_3265_0, tensor<fp16, [1, 8, 128, 8]> var_3265_1 = split(axis = var_3265_axis_0, split_sizes = var_3265_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3265")];
            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_3252_cast_fp16_0, y = var_3265_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3268_to_fp16 = const()[name = string("op_3268_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_3268_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_3272 = const()[name = string("op_3272"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_135_cast_fp16 = softmax(axis = var_3272, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3278_transpose_x_1 = const()[name = string("op_3278_transpose_x_1"), val = bool(true)];
            bool var_3278_transpose_y_1 = const()[name = string("op_3278_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_3278_cast_fp16 = matmul(transpose_x = var_3278_transpose_x_1, transpose_y = var_3278_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3262_cast_fp16_0)[name = string("op_3278_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_3252_cast_fp16_1, y = var_3265_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3280_to_fp16 = const()[name = string("op_3280_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_3280_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_3284 = const()[name = string("op_3284"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_143_cast_fp16 = softmax(axis = var_3284, x = attn_weights_141_cast_fp16)[name = string("attn_weights_143_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_143_cast_fp16, y = var_3262_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3292 = const()[name = string("op_3292"), 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_3292, interleave = attn_output_67_interleave_0, values = (var_3278_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3296_perm_0 = const()[name = string("op_3296_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_3296_cast_fp16 = transpose(perm = var_3296_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3296_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            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_cast_fp16, x = attn_output_71_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_3329_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3329_cast_fp16")];
            int32 var_3327 = const()[name = string("op_3327"), 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_3327, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3329_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(862601728)))];
            fp16 var_3339_to_fp16 = const()[name = string("op_3339_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_3339_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3350_split_sizes_0 = const()[name = string("op_3350_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3350_axis_0 = const()[name = string("op_3350_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3350_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3350_cast_fp16_1 = split(axis = var_3350_axis_0, split_sizes = var_3350_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3350_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_3367_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3367_cast_fp16")];
            tensor<int32, [2]> var_3373_strides_0 = const()[name = string("op_3373_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3373_pad_type_0 = const()[name = string("op_3373_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3373_pad_0 = const()[name = string("op_3373_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3373_dilations_0 = const()[name = string("op_3373_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3373_groups_0 = const()[name = string("op_3373_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_3373_cast_fp16 = conv(dilations = var_3373_dilations_0, groups = var_3373_groups_0, pad = var_3373_pad_0, pad_type = var_3373_pad_type_0, strides = var_3373_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("op_3373_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_89_cast_fp16 = mul(x = var_3367_cast_fp16, y = var_3373_cast_fp16)[name = string("x_89_cast_fp16")];
            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_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_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_3391_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3391_cast_fp16")];
            int32 var_3389 = const()[name = string("op_3389"), 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_3389, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3391_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_37_gamma_0_to_fp16 = const()[name = string("out_37_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862609984)))];
            fp16 var_3401_to_fp16 = const()[name = string("op_3401_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_3401_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3412_split_sizes_0 = const()[name = string("op_3412_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3412_axis_0 = const()[name = string("op_3412_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3412_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3412_cast_fp16_1 = split(axis = var_3412_axis_0, split_sizes = var_3412_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3412_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862618240)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(871006912)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3412_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_3469_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3469_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_3476_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3480_cast_fp16 = mul(x = x_91_cast_fp16, y = var_452_cast_fp16)[name = string("op_3480_cast_fp16")];
            tensor<int32, [2]> var_3481_split_sizes_0 = const()[name = string("op_3481_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3481_axis_0 = const()[name = string("op_3481_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_3481_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_3481_cast_fp16_1 = split(axis = var_3481_axis_0, split_sizes = var_3481_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3481_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_3483_cast_fp16 = mul(x = var_3481_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3483_cast_fp16")];
            int32 var_3485 = const()[name = string("op_3485"), val = int32(-2)];
            bool var_3486_interleave_0 = const()[name = string("op_3486_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_3486_cast_fp16 = concat(axis = var_3485, interleave = var_3486_interleave_0, values = (var_3483_cast_fp16, var_3481_cast_fp16_0))[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3487_cast_fp16 = mul(x = var_3486_cast_fp16, y = var_459_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_57_cast_fp16 = add(x = var_3480_cast_fp16, y = var_3487_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3493_cast_fp16 = mul(x = var_3469_cast_fp16, y = var_452_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [2]> var_3494_split_sizes_0 = const()[name = string("op_3494_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3494_axis_0 = const()[name = string("op_3494_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_3494_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_3494_cast_fp16_1 = split(axis = var_3494_axis_0, split_sizes = var_3494_split_sizes_0, x = var_3469_cast_fp16)[name = string("op_3494_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_3496_cast_fp16 = mul(x = var_3494_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3496_cast_fp16")];
            int32 var_3498 = const()[name = string("op_3498"), val = int32(-2)];
            bool var_3499_interleave_0 = const()[name = string("op_3499_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_3499_cast_fp16 = concat(axis = var_3498, interleave = var_3499_interleave_0, values = (var_3496_cast_fp16, var_3494_cast_fp16_0))[name = string("op_3499_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3500_cast_fp16 = mul(x = var_3499_cast_fp16, y = var_459_cast_fp16)[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_95_cast_fp16 = add(x = var_3493_cast_fp16, y = var_3500_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_72)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_74_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_74 = read_state(input = key_cache)[name = string("coreml_update_state_74")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3476_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_73)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_75_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_75 = read_state(input = value_cache)[name = string("coreml_update_state_75")];
            tensor<int32, [4]> var_3570_begin_0 = const()[name = string("op_3570_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3570_end_0 = const()[name = string("op_3570_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3570_end_mask_0 = const()[name = string("op_3570_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = coreml_update_state_74)[name = string("op_3570_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3573_axis_0 = const()[name = string("op_3573_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_1 = split(axis = var_3573_axis_0, split_sizes = tile_18, x = var_3570_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [4]> var_3580_begin_0 = const()[name = string("op_3580_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3580_end_0 = const()[name = string("op_3580_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3580_end_mask_0 = const()[name = string("op_3580_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3580_cast_fp16 = slice_by_index(begin = var_3580_begin_0, end = var_3580_end_0, end_mask = var_3580_end_mask_0, x = coreml_update_state_75)[name = string("op_3580_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3583_axis_0 = const()[name = string("op_3583_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_1 = split(axis = var_3583_axis_0, split_sizes = tile_19, x = var_3580_cast_fp16)[name = string("op_3583_cast_fp16")];
            tensor<int32, [2]> var_3586_split_sizes_0 = const()[name = string("op_3586_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3586_axis_0 = const()[name = string("op_3586_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_3586_0, tensor<fp16, [1, 8, 128, 8]> var_3586_1 = split(axis = var_3586_axis_0, split_sizes = var_3586_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3586")];
            bool attn_weights_145_transpose_x_0 = const()[name = string("attn_weights_145_transpose_x_0"), val = bool(false)];
            bool attn_weights_145_transpose_y_0 = const()[name = string("attn_weights_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_145_cast_fp16 = matmul(transpose_x = attn_weights_145_transpose_x_0, transpose_y = attn_weights_145_transpose_y_0, x = var_3573_cast_fp16_0, y = var_3586_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3589_to_fp16 = const()[name = string("op_3589_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_147_cast_fp16 = mul(x = attn_weights_145_cast_fp16, y = var_3589_to_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_149_cast_fp16 = add(x = attn_weights_147_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_149_cast_fp16")];
            int32 var_3593 = const()[name = string("op_3593"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_151_cast_fp16 = softmax(axis = var_3593, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3599_transpose_x_1 = const()[name = string("op_3599_transpose_x_1"), val = bool(true)];
            bool var_3599_transpose_y_1 = const()[name = string("op_3599_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_3599_cast_fp16 = matmul(transpose_x = var_3599_transpose_x_1, transpose_y = var_3599_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3583_cast_fp16_0)[name = string("op_3599_cast_fp16")];
            bool attn_weights_153_transpose_x_0 = const()[name = string("attn_weights_153_transpose_x_0"), val = bool(false)];
            bool attn_weights_153_transpose_y_0 = const()[name = string("attn_weights_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_153_cast_fp16 = matmul(transpose_x = attn_weights_153_transpose_x_0, transpose_y = attn_weights_153_transpose_y_0, x = var_3573_cast_fp16_1, y = var_3586_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3601_to_fp16 = const()[name = string("op_3601_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_155_cast_fp16 = mul(x = attn_weights_153_cast_fp16, y = var_3601_to_fp16)[name = string("attn_weights_155_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_157_cast_fp16 = add(x = attn_weights_155_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_157_cast_fp16")];
            int32 var_3605 = const()[name = string("op_3605"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_159_cast_fp16 = softmax(axis = var_3605, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3583_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3613 = const()[name = string("op_3613"), val = int32(1)];
            bool attn_output_75_interleave_0 = const()[name = string("attn_output_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_75_cast_fp16 = concat(axis = var_3613, interleave = attn_output_75_interleave_0, values = (var_3599_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3617_perm_0 = const()[name = string("op_3617_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_3617_cast_fp16 = transpose(perm = var_3617_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3617_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3650_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3650_cast_fp16")];
            int32 var_3648 = const()[name = string("op_3648"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_77_cast_fp16 = concat(axis = var_3648, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3650_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872055552)))];
            fp16 var_3660_to_fp16 = const()[name = string("op_3660_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3660_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3671_split_sizes_0 = const()[name = string("op_3671_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3671_axis_0 = const()[name = string("op_3671_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3671_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3671_cast_fp16_1 = split(axis = var_3671_axis_0, split_sizes = var_3671_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3671_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_3688_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3688_cast_fp16")];
            tensor<int32, [2]> var_3694_strides_0 = const()[name = string("op_3694_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3694_pad_type_0 = const()[name = string("op_3694_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3694_pad_0 = const()[name = string("op_3694_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3694_dilations_0 = const()[name = string("op_3694_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3694_groups_0 = const()[name = string("op_3694_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_3694_cast_fp16 = conv(dilations = var_3694_dilations_0, groups = var_3694_groups_0, pad = var_3694_pad_0, pad_type = var_3694_pad_type_0, strides = var_3694_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("op_3694_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_99_cast_fp16 = mul(x = var_3688_cast_fp16, y = var_3694_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3712_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3712_cast_fp16")];
            int32 var_3710 = const()[name = string("op_3710"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_81_cast_fp16 = concat(axis = var_3710, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3712_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872063808)))];
            fp16 var_3722_to_fp16 = const()[name = string("op_3722_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3722_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3733_split_sizes_0 = const()[name = string("op_3733_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3733_axis_0 = const()[name = string("op_3733_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3733_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3733_cast_fp16_1 = split(axis = var_3733_axis_0, split_sizes = var_3733_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872072064)))];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(880460736)))];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_cast_fp16, x = var_3733_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_3790_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3790_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_3797_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3797_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3801_cast_fp16 = mul(x = x_101_cast_fp16, y = var_452_cast_fp16)[name = string("op_3801_cast_fp16")];
            tensor<int32, [2]> var_3802_split_sizes_0 = const()[name = string("op_3802_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3802_axis_0 = const()[name = string("op_3802_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_3802_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_3802_cast_fp16_1 = split(axis = var_3802_axis_0, split_sizes = var_3802_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3802_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_3804_cast_fp16 = mul(x = var_3802_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3804_cast_fp16")];
            int32 var_3806 = const()[name = string("op_3806"), val = int32(-2)];
            bool var_3807_interleave_0 = const()[name = string("op_3807_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_3807_cast_fp16 = concat(axis = var_3806, interleave = var_3807_interleave_0, values = (var_3804_cast_fp16, var_3802_cast_fp16_0))[name = string("op_3807_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3808_cast_fp16 = mul(x = var_3807_cast_fp16, y = var_459_cast_fp16)[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_63_cast_fp16 = add(x = var_3801_cast_fp16, y = var_3808_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3814_cast_fp16 = mul(x = var_3790_cast_fp16, y = var_452_cast_fp16)[name = string("op_3814_cast_fp16")];
            tensor<int32, [2]> var_3815_split_sizes_0 = const()[name = string("op_3815_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3815_axis_0 = const()[name = string("op_3815_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_3815_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_3815_cast_fp16_1 = split(axis = var_3815_axis_0, split_sizes = var_3815_split_sizes_0, x = var_3790_cast_fp16)[name = string("op_3815_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_3817_cast_fp16 = mul(x = var_3815_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3817_cast_fp16")];
            int32 var_3819 = const()[name = string("op_3819"), val = int32(-2)];
            bool var_3820_interleave_0 = const()[name = string("op_3820_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_3820_cast_fp16 = concat(axis = var_3819, interleave = var_3820_interleave_0, values = (var_3817_cast_fp16, var_3815_cast_fp16_0))[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3821_cast_fp16 = mul(x = var_3820_cast_fp16, y = var_459_cast_fp16)[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_105_cast_fp16 = add(x = var_3814_cast_fp16, y = var_3821_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_74)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_76_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_76 = read_state(input = key_cache)[name = string("coreml_update_state_76")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3797_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_75)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_77_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_77 = read_state(input = value_cache)[name = string("coreml_update_state_77")];
            tensor<int32, [4]> var_3891_begin_0 = const()[name = string("op_3891_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3891_end_0 = const()[name = string("op_3891_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3891_end_mask_0 = const()[name = string("op_3891_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3891_cast_fp16 = slice_by_index(begin = var_3891_begin_0, end = var_3891_end_0, end_mask = var_3891_end_mask_0, x = coreml_update_state_76)[name = string("op_3891_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3894_axis_0 = const()[name = string("op_3894_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_1 = split(axis = var_3894_axis_0, split_sizes = tile_20, x = var_3891_cast_fp16)[name = string("op_3894_cast_fp16")];
            tensor<int32, [4]> var_3901_begin_0 = const()[name = string("op_3901_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3901_end_0 = const()[name = string("op_3901_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3901_end_mask_0 = const()[name = string("op_3901_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3901_cast_fp16 = slice_by_index(begin = var_3901_begin_0, end = var_3901_end_0, end_mask = var_3901_end_mask_0, x = coreml_update_state_77)[name = string("op_3901_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3904_axis_0 = const()[name = string("op_3904_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_1 = split(axis = var_3904_axis_0, split_sizes = tile_21, x = var_3901_cast_fp16)[name = string("op_3904_cast_fp16")];
            tensor<int32, [2]> var_3907_split_sizes_0 = const()[name = string("op_3907_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3907_axis_0 = const()[name = string("op_3907_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_3907_0, tensor<fp16, [1, 8, 128, 8]> var_3907_1 = split(axis = var_3907_axis_0, split_sizes = var_3907_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3907")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3894_cast_fp16_0, y = var_3907_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3910_to_fp16 = const()[name = string("op_3910_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3910_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3914 = const()[name = string("op_3914"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_167_cast_fp16 = softmax(axis = var_3914, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3920_transpose_x_1 = const()[name = string("op_3920_transpose_x_1"), val = bool(true)];
            bool var_3920_transpose_y_1 = const()[name = string("op_3920_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_3920_cast_fp16 = matmul(transpose_x = var_3920_transpose_x_1, transpose_y = var_3920_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3904_cast_fp16_0)[name = string("op_3920_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3894_cast_fp16_1, y = var_3907_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3922_to_fp16 = const()[name = string("op_3922_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3922_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3926 = const()[name = string("op_3926"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_175_cast_fp16 = softmax(axis = var_3926, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3904_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3934 = const()[name = string("op_3934"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_83_cast_fp16 = concat(axis = var_3934, interleave = attn_output_83_interleave_0, values = (var_3920_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3938_perm_0 = const()[name = string("op_3938_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_3938_cast_fp16 = transpose(perm = var_3938_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3938_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3971_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3971_cast_fp16")];
            int32 var_3969 = const()[name = string("op_3969"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_85_cast_fp16 = concat(axis = var_3969, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3971_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881509376)))];
            fp16 var_3981_to_fp16 = const()[name = string("op_3981_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3981_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3992_split_sizes_0 = const()[name = string("op_3992_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3992_axis_0 = const()[name = string("op_3992_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3992_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3992_cast_fp16_1 = split(axis = var_3992_axis_0, split_sizes = var_3992_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3992_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_4009_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [2]> var_4015_strides_0 = const()[name = string("op_4015_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4015_pad_type_0 = const()[name = string("op_4015_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4015_pad_0 = const()[name = string("op_4015_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4015_dilations_0 = const()[name = string("op_4015_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4015_groups_0 = const()[name = string("op_4015_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_4015_cast_fp16 = conv(dilations = var_4015_dilations_0, groups = var_4015_groups_0, pad = var_4015_pad_0, pad_type = var_4015_pad_type_0, strides = var_4015_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("op_4015_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_109_cast_fp16 = mul(x = var_4009_cast_fp16, y = var_4015_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4033_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4033_cast_fp16")];
            int32 var_4031 = const()[name = string("op_4031"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_89_cast_fp16 = concat(axis = var_4031, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4033_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881517632)))];
            fp16 var_4043_to_fp16 = const()[name = string("op_4043_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4043_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4054_split_sizes_0 = const()[name = string("op_4054_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4054_axis_0 = const()[name = string("op_4054_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4054_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4054_cast_fp16_1 = split(axis = var_4054_axis_0, split_sizes = var_4054_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_4111_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4111_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_4118_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4118_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4122_cast_fp16 = mul(x = x_111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4122_cast_fp16")];
            tensor<int32, [2]> var_4123_split_sizes_0 = const()[name = string("op_4123_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4123_axis_0 = const()[name = string("op_4123_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_4123_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_4123_cast_fp16_1 = split(axis = var_4123_axis_0, split_sizes = var_4123_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4123_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_4125_cast_fp16 = mul(x = var_4123_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4125_cast_fp16")];
            int32 var_4127 = const()[name = string("op_4127"), val = int32(-2)];
            bool var_4128_interleave_0 = const()[name = string("op_4128_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_4128_cast_fp16 = concat(axis = var_4127, interleave = var_4128_interleave_0, values = (var_4125_cast_fp16, var_4123_cast_fp16_0))[name = string("op_4128_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4129_cast_fp16 = mul(x = var_4128_cast_fp16, y = var_459_cast_fp16)[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_69_cast_fp16 = add(x = var_4122_cast_fp16, y = var_4129_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4135_cast_fp16 = mul(x = var_4111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4135_cast_fp16")];
            tensor<int32, [2]> var_4136_split_sizes_0 = const()[name = string("op_4136_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4136_axis_0 = const()[name = string("op_4136_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_4136_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_4136_cast_fp16_1 = split(axis = var_4136_axis_0, split_sizes = var_4136_split_sizes_0, x = var_4111_cast_fp16)[name = string("op_4136_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_4138_cast_fp16 = mul(x = var_4136_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4138_cast_fp16")];
            int32 var_4140 = const()[name = string("op_4140"), val = int32(-2)];
            bool var_4141_interleave_0 = const()[name = string("op_4141_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_4141_cast_fp16 = concat(axis = var_4140, interleave = var_4141_interleave_0, values = (var_4138_cast_fp16, var_4136_cast_fp16_0))[name = string("op_4141_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4142_cast_fp16 = mul(x = var_4141_cast_fp16, y = var_459_cast_fp16)[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_115_cast_fp16 = add(x = var_4135_cast_fp16, y = var_4142_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_76)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_78_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_78 = read_state(input = key_cache)[name = string("coreml_update_state_78")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4118_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_77)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_79_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_79 = read_state(input = value_cache)[name = string("coreml_update_state_79")];
            tensor<int32, [4]> var_4212_begin_0 = const()[name = string("op_4212_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4212_end_0 = const()[name = string("op_4212_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4212_end_mask_0 = const()[name = string("op_4212_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4212_cast_fp16 = slice_by_index(begin = var_4212_begin_0, end = var_4212_end_0, end_mask = var_4212_end_mask_0, x = coreml_update_state_78)[name = string("op_4212_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4215_axis_0 = const()[name = string("op_4215_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_1 = split(axis = var_4215_axis_0, split_sizes = tile_22, x = var_4212_cast_fp16)[name = string("op_4215_cast_fp16")];
            tensor<int32, [4]> var_4222_begin_0 = const()[name = string("op_4222_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4222_end_0 = const()[name = string("op_4222_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4222_end_mask_0 = const()[name = string("op_4222_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4222_cast_fp16 = slice_by_index(begin = var_4222_begin_0, end = var_4222_end_0, end_mask = var_4222_end_mask_0, x = coreml_update_state_79)[name = string("op_4222_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4225_axis_0 = const()[name = string("op_4225_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_1 = split(axis = var_4225_axis_0, split_sizes = tile_23, x = var_4222_cast_fp16)[name = string("op_4225_cast_fp16")];
            tensor<int32, [2]> var_4228_split_sizes_0 = const()[name = string("op_4228_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4228_axis_0 = const()[name = string("op_4228_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_4228_0, tensor<fp16, [1, 8, 128, 8]> var_4228_1 = split(axis = var_4228_axis_0, split_sizes = var_4228_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4228")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4215_cast_fp16_0, y = var_4228_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4231_to_fp16 = const()[name = string("op_4231_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4231_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4235 = const()[name = string("op_4235"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_183_cast_fp16 = softmax(axis = var_4235, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4241_transpose_x_1 = const()[name = string("op_4241_transpose_x_1"), val = bool(true)];
            bool var_4241_transpose_y_1 = const()[name = string("op_4241_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_4241_cast_fp16 = matmul(transpose_x = var_4241_transpose_x_1, transpose_y = var_4241_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4225_cast_fp16_0)[name = string("op_4241_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4215_cast_fp16_1, y = var_4228_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4243_to_fp16 = const()[name = string("op_4243_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4243_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4247 = const()[name = string("op_4247"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_191_cast_fp16 = softmax(axis = var_4247, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4225_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4255 = const()[name = string("op_4255"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_91_cast_fp16 = concat(axis = var_4255, interleave = attn_output_91_interleave_0, values = (var_4241_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4259_perm_0 = const()[name = string("op_4259_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_4259_cast_fp16 = transpose(perm = var_4259_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4259_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4292_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4292_cast_fp16")];
            int32 var_4290 = const()[name = string("op_4290"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_93_cast_fp16 = concat(axis = var_4290, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4292_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881525888)))];
            fp16 var_4302_to_fp16 = const()[name = string("op_4302_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4302_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4313_split_sizes_0 = const()[name = string("op_4313_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4313_axis_0 = const()[name = string("op_4313_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4313_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4313_cast_fp16_1 = split(axis = var_4313_axis_0, split_sizes = var_4313_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4313_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_4330_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4330_cast_fp16")];
            tensor<int32, [2]> var_4336_strides_0 = const()[name = string("op_4336_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4336_pad_type_0 = const()[name = string("op_4336_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4336_pad_0 = const()[name = string("op_4336_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4336_dilations_0 = const()[name = string("op_4336_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4336_groups_0 = const()[name = string("op_4336_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_4336_cast_fp16 = conv(dilations = var_4336_dilations_0, groups = var_4336_groups_0, pad = var_4336_pad_0, pad_type = var_4336_pad_type_0, strides = var_4336_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("op_4336_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_119_cast_fp16 = mul(x = var_4330_cast_fp16, y = var_4336_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_cast_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4354_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4354_cast_fp16")];
            int32 var_4352 = const()[name = string("op_4352"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_97_cast_fp16 = concat(axis = var_4352, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4354_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881534144)))];
            fp16 var_4364_to_fp16 = const()[name = string("op_4364_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4364_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4375_split_sizes_0 = const()[name = string("op_4375_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4375_axis_0 = const()[name = string("op_4375_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4375_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4375_cast_fp16_1 = split(axis = var_4375_axis_0, split_sizes = var_4375_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4375_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_4432_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4432_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_4439_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4439_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4443_cast_fp16 = mul(x = x_121_cast_fp16, y = var_452_cast_fp16)[name = string("op_4443_cast_fp16")];
            tensor<int32, [2]> var_4444_split_sizes_0 = const()[name = string("op_4444_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4444_axis_0 = const()[name = string("op_4444_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_4444_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_4444_cast_fp16_1 = split(axis = var_4444_axis_0, split_sizes = var_4444_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4444_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_4446_cast_fp16 = mul(x = var_4444_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4446_cast_fp16")];
            int32 var_4448 = const()[name = string("op_4448"), val = int32(-2)];
            bool var_4449_interleave_0 = const()[name = string("op_4449_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_4449_cast_fp16 = concat(axis = var_4448, interleave = var_4449_interleave_0, values = (var_4446_cast_fp16, var_4444_cast_fp16_0))[name = string("op_4449_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4450_cast_fp16 = mul(x = var_4449_cast_fp16, y = var_459_cast_fp16)[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_75_cast_fp16 = add(x = var_4443_cast_fp16, y = var_4450_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4456_cast_fp16 = mul(x = var_4432_cast_fp16, y = var_452_cast_fp16)[name = string("op_4456_cast_fp16")];
            tensor<int32, [2]> var_4457_split_sizes_0 = const()[name = string("op_4457_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4457_axis_0 = const()[name = string("op_4457_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_4457_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_4457_cast_fp16_1 = split(axis = var_4457_axis_0, split_sizes = var_4457_split_sizes_0, x = var_4432_cast_fp16)[name = string("op_4457_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_4459_cast_fp16 = mul(x = var_4457_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4459_cast_fp16")];
            int32 var_4461 = const()[name = string("op_4461"), val = int32(-2)];
            bool var_4462_interleave_0 = const()[name = string("op_4462_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_4462_cast_fp16 = concat(axis = var_4461, interleave = var_4462_interleave_0, values = (var_4459_cast_fp16, var_4457_cast_fp16_0))[name = string("op_4462_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4463_cast_fp16 = mul(x = var_4462_cast_fp16, y = var_459_cast_fp16)[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_125_cast_fp16 = add(x = var_4456_cast_fp16, y = var_4463_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_78)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_80_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_80 = read_state(input = key_cache)[name = string("coreml_update_state_80")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4439_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_79)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_81_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_81 = read_state(input = value_cache)[name = string("coreml_update_state_81")];
            tensor<int32, [4]> var_4533_begin_0 = const()[name = string("op_4533_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4533_end_0 = const()[name = string("op_4533_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4533_end_mask_0 = const()[name = string("op_4533_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4533_cast_fp16 = slice_by_index(begin = var_4533_begin_0, end = var_4533_end_0, end_mask = var_4533_end_mask_0, x = coreml_update_state_80)[name = string("op_4533_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4536_axis_0 = const()[name = string("op_4536_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_1 = split(axis = var_4536_axis_0, split_sizes = tile_24, x = var_4533_cast_fp16)[name = string("op_4536_cast_fp16")];
            tensor<int32, [4]> var_4543_begin_0 = const()[name = string("op_4543_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4543_end_0 = const()[name = string("op_4543_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4543_end_mask_0 = const()[name = string("op_4543_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4543_cast_fp16 = slice_by_index(begin = var_4543_begin_0, end = var_4543_end_0, end_mask = var_4543_end_mask_0, x = coreml_update_state_81)[name = string("op_4543_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4546_axis_0 = const()[name = string("op_4546_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_1 = split(axis = var_4546_axis_0, split_sizes = tile_25, x = var_4543_cast_fp16)[name = string("op_4546_cast_fp16")];
            tensor<int32, [2]> var_4549_split_sizes_0 = const()[name = string("op_4549_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4549_axis_0 = const()[name = string("op_4549_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_4549_0, tensor<fp16, [1, 8, 128, 8]> var_4549_1 = split(axis = var_4549_axis_0, split_sizes = var_4549_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4549")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4536_cast_fp16_0, y = var_4549_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4552_to_fp16 = const()[name = string("op_4552_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4552_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4556 = const()[name = string("op_4556"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_199_cast_fp16 = softmax(axis = var_4556, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4562_transpose_x_1 = const()[name = string("op_4562_transpose_x_1"), val = bool(true)];
            bool var_4562_transpose_y_1 = const()[name = string("op_4562_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_4562_cast_fp16 = matmul(transpose_x = var_4562_transpose_x_1, transpose_y = var_4562_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4546_cast_fp16_0)[name = string("op_4562_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4536_cast_fp16_1, y = var_4549_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4564_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4568 = const()[name = string("op_4568"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_207_cast_fp16 = softmax(axis = var_4568, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4546_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4576 = const()[name = string("op_4576"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_99_cast_fp16 = concat(axis = var_4576, interleave = attn_output_99_interleave_0, values = (var_4562_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4580_perm_0 = const()[name = string("op_4580_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_4580_cast_fp16 = transpose(perm = var_4580_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4580_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4613_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4613_cast_fp16")];
            int32 var_4611 = const()[name = string("op_4611"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_101_cast_fp16 = concat(axis = var_4611, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4613_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881542400)))];
            fp16 var_4623_to_fp16 = const()[name = string("op_4623_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4623_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4634_split_sizes_0 = const()[name = string("op_4634_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4634_axis_0 = const()[name = string("op_4634_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4634_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4634_cast_fp16_1 = split(axis = var_4634_axis_0, split_sizes = var_4634_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4634_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_4651_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4651_cast_fp16")];
            tensor<int32, [2]> var_4657_strides_0 = const()[name = string("op_4657_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4657_pad_type_0 = const()[name = string("op_4657_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4657_pad_0 = const()[name = string("op_4657_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4657_dilations_0 = const()[name = string("op_4657_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4657_groups_0 = const()[name = string("op_4657_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_4657_cast_fp16 = conv(dilations = var_4657_dilations_0, groups = var_4657_groups_0, pad = var_4657_pad_0, pad_type = var_4657_pad_type_0, strides = var_4657_strides_0, weight = layers_12_mlp_up_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("op_4657_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_129_cast_fp16 = mul(x = var_4651_cast_fp16, y = var_4657_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_cast_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4675_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4675_cast_fp16")];
            int32 var_4673 = const()[name = string("op_4673"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_105_cast_fp16 = concat(axis = var_4673, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4675_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881550656)))];
            fp16 var_4685_to_fp16 = const()[name = string("op_4685_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4685_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4696_split_sizes_0 = const()[name = string("op_4696_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4696_axis_0 = const()[name = string("op_4696_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4696_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4696_cast_fp16_1 = split(axis = var_4696_axis_0, split_sizes = var_4696_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4696_cast_fp16")];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 8]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_4753_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4753_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 8]> var_4760_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4760_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4764_cast_fp16 = mul(x = x_131_cast_fp16, y = var_452_cast_fp16)[name = string("op_4764_cast_fp16")];
            tensor<int32, [2]> var_4765_split_sizes_0 = const()[name = string("op_4765_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4765_axis_0 = const()[name = string("op_4765_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_4765_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_4765_cast_fp16_1 = split(axis = var_4765_axis_0, split_sizes = var_4765_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4765_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_4767_cast_fp16 = mul(x = var_4765_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4767_cast_fp16")];
            int32 var_4769 = const()[name = string("op_4769"), val = int32(-2)];
            bool var_4770_interleave_0 = const()[name = string("op_4770_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_4770_cast_fp16 = concat(axis = var_4769, interleave = var_4770_interleave_0, values = (var_4767_cast_fp16, var_4765_cast_fp16_0))[name = string("op_4770_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4771_cast_fp16 = mul(x = var_4770_cast_fp16, y = var_459_cast_fp16)[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_81_cast_fp16 = add(x = var_4764_cast_fp16, y = var_4771_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4777_cast_fp16 = mul(x = var_4753_cast_fp16, y = var_452_cast_fp16)[name = string("op_4777_cast_fp16")];
            tensor<int32, [2]> var_4778_split_sizes_0 = const()[name = string("op_4778_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4778_axis_0 = const()[name = string("op_4778_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_4778_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_4778_cast_fp16_1 = split(axis = var_4778_axis_0, split_sizes = var_4778_split_sizes_0, x = var_4753_cast_fp16)[name = string("op_4778_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_4780_cast_fp16 = mul(x = var_4778_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4780_cast_fp16")];
            int32 var_4782 = const()[name = string("op_4782"), val = int32(-2)];
            bool var_4783_interleave_0 = const()[name = string("op_4783_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_4783_cast_fp16 = concat(axis = var_4782, interleave = var_4783_interleave_0, values = (var_4780_cast_fp16, var_4778_cast_fp16_0))[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4784_cast_fp16 = mul(x = var_4783_cast_fp16, y = var_459_cast_fp16)[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_135_cast_fp16 = add(x = var_4777_cast_fp16, y = var_4784_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_80)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_82_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_82 = read_state(input = key_cache)[name = string("coreml_update_state_82")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4760_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_81)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_83_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_83 = read_state(input = value_cache)[name = string("coreml_update_state_83")];
            tensor<int32, [4]> var_4854_begin_0 = const()[name = string("op_4854_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4854_end_0 = const()[name = string("op_4854_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4854_end_mask_0 = const()[name = string("op_4854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4854_cast_fp16 = slice_by_index(begin = var_4854_begin_0, end = var_4854_end_0, end_mask = var_4854_end_mask_0, x = coreml_update_state_82)[name = string("op_4854_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4857_axis_0 = const()[name = string("op_4857_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_1 = split(axis = var_4857_axis_0, split_sizes = tile_26, x = var_4854_cast_fp16)[name = string("op_4857_cast_fp16")];
            tensor<int32, [4]> var_4864_begin_0 = const()[name = string("op_4864_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4864_end_0 = const()[name = string("op_4864_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4864_end_mask_0 = const()[name = string("op_4864_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4864_cast_fp16 = slice_by_index(begin = var_4864_begin_0, end = var_4864_end_0, end_mask = var_4864_end_mask_0, x = coreml_update_state_83)[name = string("op_4864_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4867_axis_0 = const()[name = string("op_4867_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_1 = split(axis = var_4867_axis_0, split_sizes = tile_27, x = var_4864_cast_fp16)[name = string("op_4867_cast_fp16")];
            tensor<int32, [2]> var_4870_split_sizes_0 = const()[name = string("op_4870_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4870_axis_0 = const()[name = string("op_4870_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_4870_0, tensor<fp16, [1, 8, 128, 8]> var_4870_1 = split(axis = var_4870_axis_0, split_sizes = var_4870_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4870")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4857_cast_fp16_0, y = var_4870_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4873_to_fp16 = const()[name = string("op_4873_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4873_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4877 = const()[name = string("op_4877"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_215_cast_fp16 = softmax(axis = var_4877, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4883_transpose_x_1 = const()[name = string("op_4883_transpose_x_1"), val = bool(true)];
            bool var_4883_transpose_y_1 = const()[name = string("op_4883_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_4883_cast_fp16 = matmul(transpose_x = var_4883_transpose_x_1, transpose_y = var_4883_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4867_cast_fp16_0)[name = string("op_4883_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4857_cast_fp16_1, y = var_4870_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4885_to_fp16 = const()[name = string("op_4885_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4885_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4889 = const()[name = string("op_4889"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_cast_fp16 = softmax(axis = var_4889, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4867_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4897 = const()[name = string("op_4897"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_107_cast_fp16 = concat(axis = var_4897, interleave = attn_output_107_interleave_0, values = (var_4883_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4901_perm_0 = const()[name = string("op_4901_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 8]> var_4901_cast_fp16 = transpose(perm = var_4901_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4901_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4934_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4934_cast_fp16")];
            int32 var_4932 = const()[name = string("op_4932"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_109_cast_fp16 = concat(axis = var_4932, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4934_cast_fp16))[name = string("doubled_109_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(881558912)))];
            fp16 var_4944_to_fp16 = const()[name = string("op_4944_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_4944_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_4955_split_sizes_0 = const()[name = string("op_4955_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4955_axis_0 = const()[name = string("op_4955_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4955_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4955_cast_fp16_1 = split(axis = var_4955_axis_0, split_sizes = var_4955_split_sizes_0, x = out_cast_fp16)[name = string("op_4955_cast_fp16")];
            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_13_mlp_gate_proj_weight_cast_fp16, x = var_4955_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_4972_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4972_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881567168)))];
            tensor<int32, [2]> var_4978_strides_0 = const()[name = string("op_4978_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4978_pad_type_0 = const()[name = string("op_4978_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4978_pad_0 = const()[name = string("op_4978_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4978_dilations_0 = const()[name = string("op_4978_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4978_groups_0 = const()[name = string("op_4978_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_4978_cast_fp16 = conv(dilations = var_4978_dilations_0, groups = var_4978_groups_0, pad = var_4978_pad_0, pad_type = var_4978_pad_type_0, strides = var_4978_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4955_cast_fp16_0)[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_cast_fp16 = mul(x = var_4972_cast_fp16, y = var_4978_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_13_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states = add(x = hidden_states_135_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_4987_cast_fp16")];
        } -> (hidden_states);
    func main<ios18>(tensor<fp32, [1, 2048, 1, 1]> inputs_embeds, state<tensor<fp16, [14, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [1]> position_index_seed, state<tensor<fp16, [14, 2, 2048, 128]>> value_cache) {
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_v_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(64))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524992))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524416))))[name = string("layers_1_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_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(525312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13120640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13108288))))[name = string("layers_1_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_v_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(13126848))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651776))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13651200))))[name = string("layers_2_self_attn_v_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(13652096))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26247424))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26235072))))[name = string("layers_2_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_v_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(26253632))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26778560))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26777984))))[name = string("layers_3_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_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(26778880))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30977408))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30973248))))[name = string("layers_3_self_attn_o_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(30979520))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43566656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43562496))))[name = string("layers_3_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_v_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(43568768))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093696))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44093120))))[name = string("layers_4_self_attn_v_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(44094016))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48292544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48288384))))[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(48294656))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60889984))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60877632))))[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(60896192))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73491520))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73479168))))[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(73497728))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86084864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86080704))))[name = string("layers_4_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_v_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(86086976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86611328))))[name = string("layers_5_self_attn_v_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(86612224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90810752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90806592))))[name = string("layers_5_self_attn_o_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(90812864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103408192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103395840))))[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(103414400))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116001536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115997376))))[name = string("layers_5_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_v_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(116003648))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528576))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116528000))))[name = string("layers_6_self_attn_v_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(116528896))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120727424))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120723264))))[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(120729536))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133324864))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133312512))))[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(133331072))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145926400))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145914048))))[name = string("layers_6_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_6_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(145932608))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158519744))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158515584))))[name = string("layers_6_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_v_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(158521856))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046784))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046208))))[name = string("layers_7_self_attn_v_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(159047104))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163245632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163241472))))[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(163247744))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175843072))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175830720))))[name = string("layers_7_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_8_self_attn_v_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(175849280))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176374208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176373632))))[name = string("layers_8_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_8_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(176374528))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180573056))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180568896))))[name = string("layers_8_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(180575168))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193170496))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193158144))))[name = string("layers_8_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_8_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(193176704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205772032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205759680))))[name = string("layers_8_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_8_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(205778240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218365376))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218361216))))[name = string("layers_8_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_v_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(218367488))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218892416))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218891840))))[name = string("layers_9_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_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(218892736))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223091264))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223087104))))[name = string("layers_9_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(223093376))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235688704))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235676352))))[name = string("layers_9_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_9_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(235694912))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248290240))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248277888))))[name = string("layers_9_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_9_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(248296448))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260883584))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260879424))))[name = string("layers_9_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_v_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(260885696))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410624))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261410048))))[name = string("layers_10_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_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(261410944))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265609472))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265605312))))[name = string("layers_10_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(265611584))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278206912))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278194560))))[name = string("layers_10_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_10_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(278213120))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290808448))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290796096))))[name = string("layers_10_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_10_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(290814656))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303401792))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303397632))))[name = string("layers_10_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(303403904))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307602432))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307598272))))[name = string("layers_11_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_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(307604544))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308129472))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308128896))))[name = string("layers_11_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_11_self_attn_v_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(308129792))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654720))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308654144))))[name = string("layers_11_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_11_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(308655040))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312853568))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312849408))))[name = string("layers_11_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(312855680))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325451008))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325438656))))[name = string("layers_11_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_11_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(325457216))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338052544))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338040192))))[name = string("layers_11_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_11_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(338058752))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350645888))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350641728))))[name = string("layers_11_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(350648000))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354846528))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354842368))))[name = string("layers_12_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_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(354848640))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355373568))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355372992))))[name = string("layers_12_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_12_self_attn_v_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(355373888))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898816))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355898240))))[name = string("layers_12_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_12_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(355899136))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360097664))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360093504))))[name = string("layers_12_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(360099776))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372695104))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372682752))))[name = string("layers_12_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_12_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(372701312))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385296640))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385284288))))[name = string("layers_12_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_12_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(385302848))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397889984))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397885824))))[name = string("layers_12_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(397892096))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402090624))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402086464))))[name = string("layers_13_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_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(402092736))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617664))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402617088))))[name = string("layers_13_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_13_self_attn_v_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(402617984))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142912))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403142336))))[name = string("layers_13_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_13_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(403143232))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407341760))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407337600))))[name = string("layers_13_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_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(407343872))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419939200))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419926848))))[name = string("layers_13_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_13_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(419945408))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432532544))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432528384))))[name = string("layers_13_mlp_down_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_424 = const()[name = string("op_424"), val = int32(0)];
            bool var_426_exclusive_0 = const()[name = string("op_426_exclusive_0"), val = bool(false)];
            bool var_426_reverse_0 = const()[name = string("op_426_reverse_0"), val = bool(false)];
            tensor<fp16, [1]> var_426_cast_fp16 = cumsum(axis = var_424, exclusive = var_426_exclusive_0, reverse = var_426_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_426_cast_fp16")];
            fp16 var_428_promoted_to_fp16 = const()[name = string("op_428_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> position_offsets_cast_fp16 = sub(x = var_426_cast_fp16, y = var_428_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_431_axes_0 = const()[name = string("op_431_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 1]> var_431_cast_fp16 = expand_dims(axes = var_431_axes_0, x = position_offsets_cast_fp16)[name = string("op_431_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_431_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(432534656)))];
            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(440923328)))];
            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_450_perm_0 = const()[name = string("op_450_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_452_axes_0 = const()[name = string("op_452_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_450_cast_fp16 = transpose(perm = var_450_perm_0, x = cos_1_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 1, 128, 1]> var_452_cast_fp16 = expand_dims(axes = var_452_axes_0, x = var_450_cast_fp16)[name = string("op_452_cast_fp16")];
            tensor<int32, [3]> var_457_perm_0 = const()[name = string("op_457_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_459_axes_0 = const()[name = string("op_459_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1]> var_457_cast_fp16 = transpose(perm = var_457_perm_0, x = sin_1_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 1, 128, 1]> var_459_cast_fp16 = expand_dims(axes = var_459_axes_0, x = var_457_cast_fp16)[name = string("op_459_cast_fp16")];
            tensor<int32, [1]> var_478_axes_0 = const()[name = string("op_478_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 1, 1]> var_478 = expand_dims(axes = var_478_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_478")];
            tensor<int32, [1, 1, 2048]> var_471 = const()[name = string("op_471"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449312000)))];
            tensor<bool, [1, 1, 2048]> var_479 = greater(x = var_471, y = var_478)[name = string("op_479")];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> var_486_axes_0 = const()[name = string("op_486_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 2048]> var_479_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_479)[name = string("cast_1")];
            tensor<fp16, [1, 1, 1, 2048]> var_486_cast_fp16 = expand_dims(axes = var_486_axes_0, x = var_479_to_fp16)[name = string("op_486_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_490_promoted_to_fp16 = const()[name = string("op_490_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 1]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_486_cast_fp16)[name = string("transpose_42")];
            tensor<bool, [1, 1, 2048, 1]> var_491_cast_fp16 = equal(x = mask_cast_fp16, y = var_490_promoted_to_fp16)[name = string("op_491_cast_fp16")];
            fp16 var_492_to_fp16 = const()[name = string("op_492_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 2048, 1]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_492_to_fp16, cond = var_491_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_502_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_2_promoted_to_fp16)[name = string("op_502_cast_fp16")];
            int32 var_500 = const()[name = string("op_500"), 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_500, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_502_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(449320256)))];
            fp16 var_512_to_fp16 = const()[name = string("op_512_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_512_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_523_split_sizes_0 = const()[name = string("op_523_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_523_axis_0 = const()[name = string("op_523_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_523_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_523_cast_fp16_1 = split(axis = var_523_axis_0, split_sizes = var_523_split_sizes_0, x = out_1_cast_fp16)[name = string("op_523_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449328512)))];
            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_to_fp16, x = var_523_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457717184)))];
            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_to_fp16, x = var_523_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(458765824)))];
            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_523_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_580_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("op_580_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_587_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("op_587_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_591_cast_fp16 = mul(x = x_1_cast_fp16, y = var_452_cast_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [2]> var_592_split_sizes_0 = const()[name = string("op_592_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_592_axis_0 = const()[name = string("op_592_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_592_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_592_cast_fp16_1 = split(axis = var_592_axis_0, split_sizes = var_592_split_sizes_0, x = x_1_cast_fp16)[name = string("op_592_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_594_cast_fp16 = mul(x = var_592_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_594_cast_fp16")];
            int32 var_596 = const()[name = string("op_596"), val = int32(-2)];
            bool var_597_interleave_0 = const()[name = string("op_597_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_597_cast_fp16 = concat(axis = var_596, interleave = var_597_interleave_0, values = (var_594_cast_fp16, var_592_cast_fp16_0))[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_598_cast_fp16 = mul(x = var_597_cast_fp16, y = var_459_cast_fp16)[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_3_cast_fp16 = add(x = var_591_cast_fp16, y = var_598_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_604_cast_fp16 = mul(x = var_580_cast_fp16, y = var_452_cast_fp16)[name = string("op_604_cast_fp16")];
            tensor<int32, [2]> var_605_split_sizes_0 = const()[name = string("op_605_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_605_axis_0 = const()[name = string("op_605_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_605_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_605_cast_fp16_1 = split(axis = var_605_axis_0, split_sizes = var_605_split_sizes_0, x = var_580_cast_fp16)[name = string("op_605_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_607_cast_fp16 = mul(x = var_605_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_607_cast_fp16")];
            int32 var_609 = const()[name = string("op_609"), val = int32(-2)];
            bool var_610_interleave_0 = const()[name = string("op_610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_610_cast_fp16 = concat(axis = var_609, interleave = var_610_interleave_0, values = (var_607_cast_fp16, var_605_cast_fp16_0))[name = string("op_610_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_611_cast_fp16 = mul(x = var_610_cast_fp16, y = var_459_cast_fp16)[name = string("op_611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_5_cast_fp16 = add(x = var_604_cast_fp16, y = var_611_cast_fp16)[name = string("key_states_5_cast_fp16")];
            tensor<fp16, [14, 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_41")];
            tensor<fp16, [14, 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, [14, 2, 2048, 128]> coreml_update_state_0 = read_state(input = key_cache)[name = string("coreml_update_state_0")];
            tensor<fp16, [14, 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_587_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [14, 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, [14, 2, 2048, 128]> coreml_update_state_1 = read_state(input = value_cache)[name = string("coreml_update_state_1")];
            tensor<int32, [4]> var_681_begin_0 = const()[name = string("op_681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_681_end_0 = const()[name = string("op_681_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_681_end_mask_0 = const()[name = string("op_681_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_681_cast_fp16 = slice_by_index(begin = var_681_begin_0, end = var_681_end_0, end_mask = var_681_end_mask_0, x = coreml_update_state_0)[name = string("op_681_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_684_axis_0 = const()[name = string("op_684_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_684_cast_fp16_1 = split(axis = var_684_axis_0, split_sizes = tile_0, x = var_681_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [4]> var_691_begin_0 = const()[name = string("op_691_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_691_end_0 = const()[name = string("op_691_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_691_end_mask_0 = const()[name = string("op_691_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = coreml_update_state_1)[name = string("op_691_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_694_axis_0 = const()[name = string("op_694_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_694_cast_fp16_1 = split(axis = var_694_axis_0, split_sizes = tile_1, x = var_691_cast_fp16)[name = string("op_694_cast_fp16")];
            tensor<int32, [2]> var_697_split_sizes_0 = const()[name = string("op_697_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_697_axis_0 = const()[name = string("op_697_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_697_0, tensor<fp16, [1, 8, 128, 1]> var_697_1 = split(axis = var_697_axis_0, split_sizes = var_697_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_697")];
            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_684_cast_fp16_0, y = var_697_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_700_to_fp16 = const()[name = string("op_700_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_700_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_704 = const()[name = string("op_704"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_7_cast_fp16 = softmax(axis = var_704, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_710_transpose_x_1 = const()[name = string("op_710_transpose_x_1"), val = bool(true)];
            bool var_710_transpose_y_1 = const()[name = string("op_710_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_710_cast_fp16 = matmul(transpose_x = var_710_transpose_x_1, transpose_y = var_710_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_694_cast_fp16_0)[name = string("op_710_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_684_cast_fp16_1, y = var_697_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_712_to_fp16 = const()[name = string("op_712_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_712_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_716 = const()[name = string("op_716"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_15_cast_fp16 = softmax(axis = var_716, 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_694_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_724 = const()[name = string("op_724"), 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_724, interleave = attn_output_3_interleave_0, values = (var_710_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_728_perm_0 = const()[name = string("op_728_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_728_cast_fp16 = transpose(perm = var_728_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_7_cast_fp16 = reshape(shape = concat_11x, x = var_728_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(459814464)))];
            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_761_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_761_cast_fp16")];
            int32 var_759 = const()[name = string("op_759"), 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_759, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_761_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(468203136)))];
            fp16 var_771_to_fp16 = const()[name = string("op_771_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_771_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_782_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = out_3_cast_fp16)[name = string("op_782_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468211392)))];
            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_to_fp16, x = var_782_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_799_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_799_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493377280)))];
            tensor<int32, [2]> var_805_strides_0 = const()[name = string("op_805_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_805_pad_type_0 = const()[name = string("op_805_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_805_pad_0 = const()[name = string("op_805_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_805_dilations_0 = const()[name = string("op_805_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_805_groups_0 = const()[name = string("op_805_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_805_cast_fp16 = conv(dilations = var_805_dilations_0, groups = var_805_groups_0, pad = var_805_pad_0, pad_type = var_805_pad_type_0, strides = var_805_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_782_cast_fp16_0)[name = string("op_805_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_9_cast_fp16 = mul(x = var_799_cast_fp16, y = var_805_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518543168)))];
            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_to_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_823_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_823_cast_fp16")];
            int32 var_821 = const()[name = string("op_821"), 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_821, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_823_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(543709056)))];
            fp16 var_833_to_fp16 = const()[name = string("op_833_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_833_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_844_split_sizes_0 = const()[name = string("op_844_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_844_axis_0 = const()[name = string("op_844_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_844_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_844_cast_fp16_1 = split(axis = var_844_axis_0, split_sizes = var_844_split_sizes_0, x = out_5_cast_fp16)[name = string("op_844_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543717312)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552105984)))];
            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_to_fp16, x = var_844_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            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_cast_fp16, x = var_844_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_901_cast_fp16 = reshape(shape = concat_13x, x = key_states_11_cast_fp16)[name = string("op_901_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_908_cast_fp16 = reshape(shape = concat_14x, x = value_states_7_cast_fp16)[name = string("op_908_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_912_cast_fp16 = mul(x = x_11_cast_fp16, y = var_452_cast_fp16)[name = string("op_912_cast_fp16")];
            tensor<int32, [2]> var_913_split_sizes_0 = const()[name = string("op_913_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_913_axis_0 = const()[name = string("op_913_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_913_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_913_cast_fp16_1 = split(axis = var_913_axis_0, split_sizes = var_913_split_sizes_0, x = x_11_cast_fp16)[name = string("op_913_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_915_cast_fp16 = mul(x = var_913_cast_fp16_1, y = const_14_promoted_to_fp16)[name = string("op_915_cast_fp16")];
            int32 var_917 = const()[name = string("op_917"), val = int32(-2)];
            bool var_918_interleave_0 = const()[name = string("op_918_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_918_cast_fp16 = concat(axis = var_917, interleave = var_918_interleave_0, values = (var_915_cast_fp16, var_913_cast_fp16_0))[name = string("op_918_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_919_cast_fp16 = mul(x = var_918_cast_fp16, y = var_459_cast_fp16)[name = string("op_919_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_9_cast_fp16 = add(x = var_912_cast_fp16, y = var_919_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_925_cast_fp16 = mul(x = var_901_cast_fp16, y = var_452_cast_fp16)[name = string("op_925_cast_fp16")];
            tensor<int32, [2]> var_926_split_sizes_0 = const()[name = string("op_926_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_926_axis_0 = const()[name = string("op_926_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_926_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_926_cast_fp16_1 = split(axis = var_926_axis_0, split_sizes = var_926_split_sizes_0, x = var_901_cast_fp16)[name = string("op_926_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_928_cast_fp16 = mul(x = var_926_cast_fp16_1, y = const_15_promoted_to_fp16)[name = string("op_928_cast_fp16")];
            int32 var_930 = const()[name = string("op_930"), val = int32(-2)];
            bool var_931_interleave_0 = const()[name = string("op_931_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_931_cast_fp16 = concat(axis = var_930, interleave = var_931_interleave_0, values = (var_928_cast_fp16, var_926_cast_fp16_0))[name = string("op_931_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_932_cast_fp16 = mul(x = var_931_cast_fp16, y = var_459_cast_fp16)[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_15_cast_fp16 = add(x = var_925_cast_fp16, y = var_932_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_38")];
            tensor<fp16, [14, 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, [14, 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_908_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [14, 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, [14, 2, 2048, 128]> coreml_update_state_3 = read_state(input = value_cache)[name = string("coreml_update_state_3")];
            tensor<int32, [4]> var_1002_begin_0 = const()[name = string("op_1002_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1002_end_0 = const()[name = string("op_1002_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1002_end_mask_0 = const()[name = string("op_1002_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1002_cast_fp16 = slice_by_index(begin = var_1002_begin_0, end = var_1002_end_0, end_mask = var_1002_end_mask_0, x = coreml_update_state_2)[name = string("op_1002_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1005_axis_0 = const()[name = string("op_1005_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1005_cast_fp16_1 = split(axis = var_1005_axis_0, split_sizes = tile_2, x = var_1002_cast_fp16)[name = string("op_1005_cast_fp16")];
            tensor<int32, [4]> var_1012_begin_0 = const()[name = string("op_1012_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1012_end_0 = const()[name = string("op_1012_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1012_end_mask_0 = const()[name = string("op_1012_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1012_cast_fp16 = slice_by_index(begin = var_1012_begin_0, end = var_1012_end_0, end_mask = var_1012_end_mask_0, x = coreml_update_state_3)[name = string("op_1012_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1015_axis_0 = const()[name = string("op_1015_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1015_cast_fp16_1 = split(axis = var_1015_axis_0, split_sizes = tile_3, x = var_1012_cast_fp16)[name = string("op_1015_cast_fp16")];
            tensor<int32, [2]> var_1018_split_sizes_0 = const()[name = string("op_1018_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1018_axis_0 = const()[name = string("op_1018_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1018_0, tensor<fp16, [1, 8, 128, 1]> var_1018_1 = split(axis = var_1018_axis_0, split_sizes = var_1018_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1018")];
            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_1005_cast_fp16_0, y = var_1018_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1021_to_fp16 = const()[name = string("op_1021_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_1021_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_1025 = const()[name = string("op_1025"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_23_cast_fp16 = softmax(axis = var_1025, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1031_transpose_x_1 = const()[name = string("op_1031_transpose_x_1"), val = bool(true)];
            bool var_1031_transpose_y_1 = const()[name = string("op_1031_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1031_cast_fp16 = matmul(transpose_x = var_1031_transpose_x_1, transpose_y = var_1031_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1015_cast_fp16_0)[name = string("op_1031_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_1005_cast_fp16_1, y = var_1018_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1033_to_fp16 = const()[name = string("op_1033_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_1033_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_1037 = const()[name = string("op_1037"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_31_cast_fp16 = softmax(axis = var_1037, 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_1015_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1045 = const()[name = string("op_1045"), 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_1045, interleave = attn_output_11_interleave_0, values = (var_1031_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1049_perm_0 = const()[name = string("op_1049_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_1049_cast_fp16 = transpose(perm = var_1049_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_15_cast_fp16 = reshape(shape = concat_23x, x = var_1049_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553154624)))];
            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_to_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_1082_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1082_cast_fp16")];
            int32 var_1080 = const()[name = string("op_1080"), 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_1080, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1082_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(561543296)))];
            fp16 var_1092_to_fp16 = const()[name = string("op_1092_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_1092_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1103_split_sizes_0 = const()[name = string("op_1103_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1103_axis_0 = const()[name = string("op_1103_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1103_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1103_cast_fp16_1 = split(axis = var_1103_axis_0, split_sizes = var_1103_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1103_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561551552)))];
            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_to_fp16, x = var_1103_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1120_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [2]> var_1126_strides_0 = const()[name = string("op_1126_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1126_pad_type_0 = const()[name = string("op_1126_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1126_pad_0 = const()[name = string("op_1126_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1126_dilations_0 = const()[name = string("op_1126_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1126_groups_0 = const()[name = string("op_1126_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1126_cast_fp16 = conv(dilations = var_1126_dilations_0, groups = var_1126_groups_0, pad = var_1126_pad_0, pad_type = var_1126_pad_type_0, strides = var_1126_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1103_cast_fp16_0)[name = string("op_1126_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_19_cast_fp16 = mul(x = var_1120_cast_fp16, y = var_1126_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(586717440)))];
            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_to_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_1144_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1144_cast_fp16")];
            int32 var_1142 = const()[name = string("op_1142"), 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_1142, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1144_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(611883328)))];
            fp16 var_1154_to_fp16 = const()[name = string("op_1154_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_1154_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1165_split_sizes_0 = const()[name = string("op_1165_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1165_axis_0 = const()[name = string("op_1165_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1165_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1165_cast_fp16_1 = split(axis = var_1165_axis_0, split_sizes = var_1165_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1165_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611891584)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620280256)))];
            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_to_fp16, x = var_1165_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            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_cast_fp16, x = var_1165_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_1222_cast_fp16 = reshape(shape = concat_25x, x = key_states_21_cast_fp16)[name = string("op_1222_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_1229_cast_fp16 = reshape(shape = concat_26x, x = value_states_13_cast_fp16)[name = string("op_1229_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1233_cast_fp16 = mul(x = x_21_cast_fp16, y = var_452_cast_fp16)[name = string("op_1233_cast_fp16")];
            tensor<int32, [2]> var_1234_split_sizes_0 = const()[name = string("op_1234_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1234_axis_0 = const()[name = string("op_1234_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1234_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1234_cast_fp16_1 = split(axis = var_1234_axis_0, split_sizes = var_1234_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1234_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_1236_cast_fp16 = mul(x = var_1234_cast_fp16_1, y = const_24_promoted_to_fp16)[name = string("op_1236_cast_fp16")];
            int32 var_1238 = const()[name = string("op_1238"), val = int32(-2)];
            bool var_1239_interleave_0 = const()[name = string("op_1239_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1239_cast_fp16 = concat(axis = var_1238, interleave = var_1239_interleave_0, values = (var_1236_cast_fp16, var_1234_cast_fp16_0))[name = string("op_1239_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1240_cast_fp16 = mul(x = var_1239_cast_fp16, y = var_459_cast_fp16)[name = string("op_1240_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_15_cast_fp16 = add(x = var_1233_cast_fp16, y = var_1240_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1246_cast_fp16 = mul(x = var_1222_cast_fp16, y = var_452_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<int32, [2]> var_1247_split_sizes_0 = const()[name = string("op_1247_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1247_axis_0 = const()[name = string("op_1247_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1247_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1247_cast_fp16_1 = split(axis = var_1247_axis_0, split_sizes = var_1247_split_sizes_0, x = var_1222_cast_fp16)[name = string("op_1247_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_1249_cast_fp16 = mul(x = var_1247_cast_fp16_1, y = const_25_promoted_to_fp16)[name = string("op_1249_cast_fp16")];
            int32 var_1251 = const()[name = string("op_1251"), val = int32(-2)];
            bool var_1252_interleave_0 = const()[name = string("op_1252_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1252_cast_fp16 = concat(axis = var_1251, interleave = var_1252_interleave_0, values = (var_1249_cast_fp16, var_1247_cast_fp16_0))[name = string("op_1252_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1253_cast_fp16 = mul(x = var_1252_cast_fp16, y = var_459_cast_fp16)[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_25_cast_fp16 = add(x = var_1246_cast_fp16, y = var_1253_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_35")];
            tensor<fp16, [14, 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, [14, 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_1229_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [14, 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, [14, 2, 2048, 128]> coreml_update_state_5 = read_state(input = value_cache)[name = string("coreml_update_state_5")];
            tensor<int32, [4]> var_1323_begin_0 = const()[name = string("op_1323_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1323_end_0 = const()[name = string("op_1323_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1323_end_mask_0 = const()[name = string("op_1323_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1323_cast_fp16 = slice_by_index(begin = var_1323_begin_0, end = var_1323_end_0, end_mask = var_1323_end_mask_0, x = coreml_update_state_4)[name = string("op_1323_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1326_axis_0 = const()[name = string("op_1326_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1326_cast_fp16_1 = split(axis = var_1326_axis_0, split_sizes = tile_4, x = var_1323_cast_fp16)[name = string("op_1326_cast_fp16")];
            tensor<int32, [4]> var_1333_begin_0 = const()[name = string("op_1333_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1333_end_0 = const()[name = string("op_1333_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1333_end_mask_0 = const()[name = string("op_1333_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1333_cast_fp16 = slice_by_index(begin = var_1333_begin_0, end = var_1333_end_0, end_mask = var_1333_end_mask_0, x = coreml_update_state_5)[name = string("op_1333_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1336_axis_0 = const()[name = string("op_1336_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1336_cast_fp16_1 = split(axis = var_1336_axis_0, split_sizes = tile_5, x = var_1333_cast_fp16)[name = string("op_1336_cast_fp16")];
            tensor<int32, [2]> var_1339_split_sizes_0 = const()[name = string("op_1339_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1339_axis_0 = const()[name = string("op_1339_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1339_0, tensor<fp16, [1, 8, 128, 1]> var_1339_1 = split(axis = var_1339_axis_0, split_sizes = var_1339_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1339")];
            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_1326_cast_fp16_0, y = var_1339_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1342_to_fp16 = const()[name = string("op_1342_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_1342_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_1346 = const()[name = string("op_1346"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_39_cast_fp16 = softmax(axis = var_1346, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1352_transpose_x_1 = const()[name = string("op_1352_transpose_x_1"), val = bool(true)];
            bool var_1352_transpose_y_1 = const()[name = string("op_1352_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1352_cast_fp16 = matmul(transpose_x = var_1352_transpose_x_1, transpose_y = var_1352_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1336_cast_fp16_0)[name = string("op_1352_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_1326_cast_fp16_1, y = var_1339_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1354_to_fp16 = const()[name = string("op_1354_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_1354_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_1358 = const()[name = string("op_1358"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_47_cast_fp16 = softmax(axis = var_1358, 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_1336_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1366 = const()[name = string("op_1366"), 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_1366, interleave = attn_output_19_interleave_0, values = (var_1352_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1370_perm_0 = const()[name = string("op_1370_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_1370_cast_fp16 = transpose(perm = var_1370_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_23_cast_fp16 = reshape(shape = concat_35x, x = var_1370_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(621328896)))];
            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_to_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_1403_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1403_cast_fp16")];
            int32 var_1401 = const()[name = string("op_1401"), 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_1401, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1403_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(629717568)))];
            fp16 var_1413_to_fp16 = const()[name = string("op_1413_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_1413_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1424_split_sizes_0 = const()[name = string("op_1424_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1424_axis_0 = const()[name = string("op_1424_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1424_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1424_cast_fp16_1 = split(axis = var_1424_axis_0, split_sizes = var_1424_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1424_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629725824)))];
            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_to_fp16, x = var_1424_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1441_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1441_cast_fp16")];
            tensor<int32, [2]> var_1447_strides_0 = const()[name = string("op_1447_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1447_pad_type_0 = const()[name = string("op_1447_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1447_pad_0 = const()[name = string("op_1447_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1447_dilations_0 = const()[name = string("op_1447_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1447_groups_0 = const()[name = string("op_1447_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1447_cast_fp16 = conv(dilations = var_1447_dilations_0, groups = var_1447_groups_0, pad = var_1447_pad_0, pad_type = var_1447_pad_type_0, strides = var_1447_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1424_cast_fp16_0)[name = string("op_1447_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_29_cast_fp16 = mul(x = var_1441_cast_fp16, y = var_1447_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654891712)))];
            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_to_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_1465_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1465_cast_fp16")];
            int32 var_1463 = const()[name = string("op_1463"), 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_1463, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1465_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(680057600)))];
            fp16 var_1475_to_fp16 = const()[name = string("op_1475_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_1475_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1486_split_sizes_0 = const()[name = string("op_1486_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1486_axis_0 = const()[name = string("op_1486_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1486_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1486_cast_fp16_1 = split(axis = var_1486_axis_0, split_sizes = var_1486_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1486_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680065856)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688454528)))];
            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_to_fp16, x = var_1486_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            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_cast_fp16, x = var_1486_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_1543_cast_fp16 = reshape(shape = concat_37x, x = key_states_31_cast_fp16)[name = string("op_1543_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_1550_cast_fp16 = reshape(shape = concat_38x, x = value_states_19_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1554_cast_fp16 = mul(x = x_31_cast_fp16, y = var_452_cast_fp16)[name = string("op_1554_cast_fp16")];
            tensor<int32, [2]> var_1555_split_sizes_0 = const()[name = string("op_1555_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1555_axis_0 = const()[name = string("op_1555_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1555_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1555_cast_fp16_1 = split(axis = var_1555_axis_0, split_sizes = var_1555_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1555_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_1557_cast_fp16 = mul(x = var_1555_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1557_cast_fp16")];
            int32 var_1559 = const()[name = string("op_1559"), val = int32(-2)];
            bool var_1560_interleave_0 = const()[name = string("op_1560_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1560_cast_fp16 = concat(axis = var_1559, interleave = var_1560_interleave_0, values = (var_1557_cast_fp16, var_1555_cast_fp16_0))[name = string("op_1560_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1561_cast_fp16 = mul(x = var_1560_cast_fp16, y = var_459_cast_fp16)[name = string("op_1561_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_21_cast_fp16 = add(x = var_1554_cast_fp16, y = var_1561_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1567_cast_fp16 = mul(x = var_1543_cast_fp16, y = var_452_cast_fp16)[name = string("op_1567_cast_fp16")];
            tensor<int32, [2]> var_1568_split_sizes_0 = const()[name = string("op_1568_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1568_axis_0 = const()[name = string("op_1568_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1568_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1568_cast_fp16_1 = split(axis = var_1568_axis_0, split_sizes = var_1568_split_sizes_0, x = var_1543_cast_fp16)[name = string("op_1568_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_1570_cast_fp16 = mul(x = var_1568_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1570_cast_fp16")];
            int32 var_1572 = const()[name = string("op_1572"), val = int32(-2)];
            bool var_1573_interleave_0 = const()[name = string("op_1573_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1573_cast_fp16 = concat(axis = var_1572, interleave = var_1573_interleave_0, values = (var_1570_cast_fp16, var_1568_cast_fp16_0))[name = string("op_1573_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1574_cast_fp16 = mul(x = var_1573_cast_fp16, y = var_459_cast_fp16)[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_35_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1574_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_32")];
            tensor<fp16, [14, 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, [14, 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_1550_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [14, 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, [14, 2, 2048, 128]> coreml_update_state_7 = read_state(input = value_cache)[name = string("coreml_update_state_7")];
            tensor<int32, [4]> var_1644_begin_0 = const()[name = string("op_1644_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1644_end_0 = const()[name = string("op_1644_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1644_end_mask_0 = const()[name = string("op_1644_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1644_cast_fp16 = slice_by_index(begin = var_1644_begin_0, end = var_1644_end_0, end_mask = var_1644_end_mask_0, x = coreml_update_state_6)[name = string("op_1644_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1647_axis_0 = const()[name = string("op_1647_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1647_cast_fp16_1 = split(axis = var_1647_axis_0, split_sizes = tile_6, x = var_1644_cast_fp16)[name = string("op_1647_cast_fp16")];
            tensor<int32, [4]> var_1654_begin_0 = const()[name = string("op_1654_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1654_end_0 = const()[name = string("op_1654_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_1654_end_mask_0 = const()[name = string("op_1654_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1654_cast_fp16 = slice_by_index(begin = var_1654_begin_0, end = var_1654_end_0, end_mask = var_1654_end_mask_0, x = coreml_update_state_7)[name = string("op_1654_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1657_axis_0 = const()[name = string("op_1657_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1657_cast_fp16_1 = split(axis = var_1657_axis_0, split_sizes = tile_7, x = var_1654_cast_fp16)[name = string("op_1657_cast_fp16")];
            tensor<int32, [2]> var_1660_split_sizes_0 = const()[name = string("op_1660_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1660_axis_0 = const()[name = string("op_1660_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1660_0, tensor<fp16, [1, 8, 128, 1]> var_1660_1 = split(axis = var_1660_axis_0, split_sizes = var_1660_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_1660")];
            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_1647_cast_fp16_0, y = var_1660_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_1663_to_fp16 = const()[name = string("op_1663_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_1663_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_1667 = const()[name = string("op_1667"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_55_cast_fp16 = softmax(axis = var_1667, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_1673_transpose_x_1 = const()[name = string("op_1673_transpose_x_1"), val = bool(true)];
            bool var_1673_transpose_y_1 = const()[name = string("op_1673_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1673_cast_fp16 = matmul(transpose_x = var_1673_transpose_x_1, transpose_y = var_1673_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_1657_cast_fp16_0)[name = string("op_1673_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_1647_cast_fp16_1, y = var_1660_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_1675_to_fp16 = const()[name = string("op_1675_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_1675_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_1679 = const()[name = string("op_1679"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_63_cast_fp16 = softmax(axis = var_1679, 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_1657_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_1687 = const()[name = string("op_1687"), 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_1687, interleave = attn_output_27_interleave_0, values = (var_1673_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_1691_perm_0 = const()[name = string("op_1691_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_1691_cast_fp16 = transpose(perm = var_1691_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_31_cast_fp16 = reshape(shape = concat_47x, x = var_1691_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            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_cast_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_1724_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1724_cast_fp16")];
            int32 var_1722 = const()[name = string("op_1722"), 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_1722, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_1724_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(689503168)))];
            fp16 var_1734_to_fp16 = const()[name = string("op_1734_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_1734_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_1745_split_sizes_0 = const()[name = string("op_1745_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1745_axis_0 = const()[name = string("op_1745_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1745_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1745_cast_fp16_1 = split(axis = var_1745_axis_0, split_sizes = var_1745_split_sizes_0, x = out_15_cast_fp16)[name = string("op_1745_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689511424)))];
            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_to_fp16, x = var_1745_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_1762_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1762_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(714677312)))];
            tensor<int32, [2]> var_1768_strides_0 = const()[name = string("op_1768_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_1768_pad_type_0 = const()[name = string("op_1768_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_1768_pad_0 = const()[name = string("op_1768_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1768_dilations_0 = const()[name = string("op_1768_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1768_groups_0 = const()[name = string("op_1768_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_1768_cast_fp16 = conv(dilations = var_1768_dilations_0, groups = var_1768_groups_0, pad = var_1768_pad_0, pad_type = var_1768_pad_type_0, strides = var_1768_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_1745_cast_fp16_0)[name = string("op_1768_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_39_cast_fp16 = mul(x = var_1762_cast_fp16, y = var_1768_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_1786_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1786_cast_fp16")];
            int32 var_1784 = const()[name = string("op_1784"), 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_1784, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_1786_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(739843200)))];
            fp16 var_1796_to_fp16 = const()[name = string("op_1796_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_1796_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_1807_split_sizes_0 = const()[name = string("op_1807_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1807_axis_0 = const()[name = string("op_1807_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_1807_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_1807_cast_fp16_1 = split(axis = var_1807_axis_0, split_sizes = var_1807_split_sizes_0, x = out_17_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739851456)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(748240128)))];
            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_to_fp16, x = var_1807_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            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_cast_fp16, x = var_1807_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_1864_cast_fp16 = reshape(shape = concat_49x, x = key_states_41_cast_fp16)[name = string("op_1864_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_1871_cast_fp16 = reshape(shape = concat_50x, x = value_states_25_cast_fp16)[name = string("op_1871_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1875_cast_fp16 = mul(x = x_41_cast_fp16, y = var_452_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [2]> var_1876_split_sizes_0 = const()[name = string("op_1876_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1876_axis_0 = const()[name = string("op_1876_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_1876_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_1876_cast_fp16_1 = split(axis = var_1876_axis_0, split_sizes = var_1876_split_sizes_0, x = x_41_cast_fp16)[name = string("op_1876_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_1878_cast_fp16 = mul(x = var_1876_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1878_cast_fp16")];
            int32 var_1880 = const()[name = string("op_1880"), val = int32(-2)];
            bool var_1881_interleave_0 = const()[name = string("op_1881_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_1881_cast_fp16 = concat(axis = var_1880, interleave = var_1881_interleave_0, values = (var_1878_cast_fp16, var_1876_cast_fp16_0))[name = string("op_1881_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_1882_cast_fp16 = mul(x = var_1881_cast_fp16, y = var_459_cast_fp16)[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_27_cast_fp16 = add(x = var_1875_cast_fp16, y = var_1882_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1888_cast_fp16 = mul(x = var_1864_cast_fp16, y = var_452_cast_fp16)[name = string("op_1888_cast_fp16")];
            tensor<int32, [2]> var_1889_split_sizes_0 = const()[name = string("op_1889_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1889_axis_0 = const()[name = string("op_1889_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_1889_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_1889_cast_fp16_1 = split(axis = var_1889_axis_0, split_sizes = var_1889_split_sizes_0, x = var_1864_cast_fp16)[name = string("op_1889_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_1891_cast_fp16 = mul(x = var_1889_cast_fp16_1, y = const_45_promoted_to_fp16)[name = string("op_1891_cast_fp16")];
            int32 var_1893 = const()[name = string("op_1893"), val = int32(-2)];
            bool var_1894_interleave_0 = const()[name = string("op_1894_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_1894_cast_fp16 = concat(axis = var_1893, interleave = var_1894_interleave_0, values = (var_1891_cast_fp16, var_1889_cast_fp16_0))[name = string("op_1894_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_1895_cast_fp16 = mul(x = var_1894_cast_fp16, y = var_459_cast_fp16)[name = string("op_1895_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_45_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1895_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_29")];
            tensor<fp16, [14, 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, [14, 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_1871_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [14, 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, [14, 2, 2048, 128]> coreml_update_state_9 = read_state(input = value_cache)[name = string("coreml_update_state_9")];
            tensor<int32, [4]> var_1965_begin_0 = const()[name = string("op_1965_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1965_end_0 = const()[name = string("op_1965_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1965_end_mask_0 = const()[name = string("op_1965_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1965_cast_fp16 = slice_by_index(begin = var_1965_begin_0, end = var_1965_end_0, end_mask = var_1965_end_mask_0, x = coreml_update_state_8)[name = string("op_1965_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1968_axis_0 = const()[name = string("op_1968_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1968_cast_fp16_1 = split(axis = var_1968_axis_0, split_sizes = tile_8, x = var_1965_cast_fp16)[name = string("op_1968_cast_fp16")];
            tensor<int32, [4]> var_1975_begin_0 = const()[name = string("op_1975_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1975_end_0 = const()[name = string("op_1975_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_1975_end_mask_0 = const()[name = string("op_1975_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1975_cast_fp16 = slice_by_index(begin = var_1975_begin_0, end = var_1975_end_0, end_mask = var_1975_end_mask_0, x = coreml_update_state_9)[name = string("op_1975_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1978_axis_0 = const()[name = string("op_1978_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1978_cast_fp16_1 = split(axis = var_1978_axis_0, split_sizes = tile_9, x = var_1975_cast_fp16)[name = string("op_1978_cast_fp16")];
            tensor<int32, [2]> var_1981_split_sizes_0 = const()[name = string("op_1981_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1981_axis_0 = const()[name = string("op_1981_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_1981_0, tensor<fp16, [1, 8, 128, 1]> var_1981_1 = split(axis = var_1981_axis_0, split_sizes = var_1981_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1981")];
            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_1968_cast_fp16_0, y = var_1981_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_1984_to_fp16 = const()[name = string("op_1984_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_1984_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_1988 = const()[name = string("op_1988"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_71_cast_fp16 = softmax(axis = var_1988, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_1994_transpose_x_1 = const()[name = string("op_1994_transpose_x_1"), val = bool(true)];
            bool var_1994_transpose_y_1 = const()[name = string("op_1994_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_1994_cast_fp16 = matmul(transpose_x = var_1994_transpose_x_1, transpose_y = var_1994_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_1978_cast_fp16_0)[name = string("op_1994_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_1968_cast_fp16_1, y = var_1981_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_1996_to_fp16 = const()[name = string("op_1996_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_1996_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_2000 = const()[name = string("op_2000"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_79_cast_fp16 = softmax(axis = var_2000, 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_1978_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2008 = const()[name = string("op_2008"), 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_2008, interleave = attn_output_35_interleave_0, values = (var_1994_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2012_perm_0 = const()[name = string("op_2012_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_2012_cast_fp16 = transpose(perm = var_2012_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_39_cast_fp16 = reshape(shape = concat_59x, x = var_2012_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_2045_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_2045_cast_fp16")];
            int32 var_2043 = const()[name = string("op_2043"), 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_2043, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2045_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(749288768)))];
            fp16 var_2055_to_fp16 = const()[name = string("op_2055_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_2055_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2066_split_sizes_0 = const()[name = string("op_2066_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2066_axis_0 = const()[name = string("op_2066_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2066_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2066_cast_fp16_1 = split(axis = var_2066_axis_0, split_sizes = var_2066_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2066_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_2066_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2083_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2083_cast_fp16")];
            tensor<int32, [2]> var_2089_strides_0 = const()[name = string("op_2089_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2089_pad_type_0 = const()[name = string("op_2089_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2089_pad_0 = const()[name = string("op_2089_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2089_dilations_0 = const()[name = string("op_2089_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2089_groups_0 = const()[name = string("op_2089_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2089_cast_fp16 = conv(dilations = var_2089_dilations_0, groups = var_2089_groups_0, pad = var_2089_pad_0, pad_type = var_2089_pad_type_0, strides = var_2089_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2066_cast_fp16_0)[name = string("op_2089_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_49_cast_fp16 = mul(x = var_2083_cast_fp16, y = var_2089_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_2107_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_2107_cast_fp16")];
            int32 var_2105 = const()[name = string("op_2105"), 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_2105, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2107_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(749297024)))];
            fp16 var_2117_to_fp16 = const()[name = string("op_2117_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_2117_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2128_split_sizes_0 = const()[name = string("op_2128_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2128_axis_0 = const()[name = string("op_2128_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2128_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2128_cast_fp16_1 = split(axis = var_2128_axis_0, split_sizes = var_2128_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749305280)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(757693952)))];
            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_to_fp16, x = var_2128_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            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_cast_fp16, x = var_2128_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_2185_cast_fp16 = reshape(shape = concat_61x, x = key_states_51_cast_fp16)[name = string("op_2185_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_2192_cast_fp16 = reshape(shape = concat_62x, x = value_states_31_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2196_cast_fp16 = mul(x = x_51_cast_fp16, y = var_452_cast_fp16)[name = string("op_2196_cast_fp16")];
            tensor<int32, [2]> var_2197_split_sizes_0 = const()[name = string("op_2197_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2197_axis_0 = const()[name = string("op_2197_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2197_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2197_cast_fp16_1 = split(axis = var_2197_axis_0, split_sizes = var_2197_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2197_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_2199_cast_fp16 = mul(x = var_2197_cast_fp16_1, y = const_54_promoted_to_fp16)[name = string("op_2199_cast_fp16")];
            int32 var_2201 = const()[name = string("op_2201"), val = int32(-2)];
            bool var_2202_interleave_0 = const()[name = string("op_2202_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2202_cast_fp16 = concat(axis = var_2201, interleave = var_2202_interleave_0, values = (var_2199_cast_fp16, var_2197_cast_fp16_0))[name = string("op_2202_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2203_cast_fp16 = mul(x = var_2202_cast_fp16, y = var_459_cast_fp16)[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_33_cast_fp16 = add(x = var_2196_cast_fp16, y = var_2203_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2209_cast_fp16 = mul(x = var_2185_cast_fp16, y = var_452_cast_fp16)[name = string("op_2209_cast_fp16")];
            tensor<int32, [2]> var_2210_split_sizes_0 = const()[name = string("op_2210_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2210_axis_0 = const()[name = string("op_2210_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2210_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2210_cast_fp16_1 = split(axis = var_2210_axis_0, split_sizes = var_2210_split_sizes_0, x = var_2185_cast_fp16)[name = string("op_2210_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_2212_cast_fp16 = mul(x = var_2210_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2212_cast_fp16")];
            int32 var_2214 = const()[name = string("op_2214"), val = int32(-2)];
            bool var_2215_interleave_0 = const()[name = string("op_2215_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2215_cast_fp16 = concat(axis = var_2214, interleave = var_2215_interleave_0, values = (var_2212_cast_fp16, var_2210_cast_fp16_0))[name = string("op_2215_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2216_cast_fp16 = mul(x = var_2215_cast_fp16, y = var_459_cast_fp16)[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_55_cast_fp16 = add(x = var_2209_cast_fp16, y = var_2216_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_26")];
            tensor<fp16, [14, 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, [14, 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_2192_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [14, 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, [14, 2, 2048, 128]> coreml_update_state_11 = read_state(input = value_cache)[name = string("coreml_update_state_11")];
            tensor<int32, [4]> var_2286_begin_0 = const()[name = string("op_2286_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2286_end_0 = const()[name = string("op_2286_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2286_end_mask_0 = const()[name = string("op_2286_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2286_cast_fp16 = slice_by_index(begin = var_2286_begin_0, end = var_2286_end_0, end_mask = var_2286_end_mask_0, x = coreml_update_state_10)[name = string("op_2286_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2289_axis_0 = const()[name = string("op_2289_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2289_cast_fp16_1 = split(axis = var_2289_axis_0, split_sizes = tile_10, x = var_2286_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> var_2296_begin_0 = const()[name = string("op_2296_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2296_end_0 = const()[name = string("op_2296_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2296_end_mask_0 = const()[name = string("op_2296_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2296_cast_fp16 = slice_by_index(begin = var_2296_begin_0, end = var_2296_end_0, end_mask = var_2296_end_mask_0, x = coreml_update_state_11)[name = string("op_2296_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2299_axis_0 = const()[name = string("op_2299_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2299_cast_fp16_1 = split(axis = var_2299_axis_0, split_sizes = tile_11, x = var_2296_cast_fp16)[name = string("op_2299_cast_fp16")];
            tensor<int32, [2]> var_2302_split_sizes_0 = const()[name = string("op_2302_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2302_axis_0 = const()[name = string("op_2302_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2302_0, tensor<fp16, [1, 8, 128, 1]> var_2302_1 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2302")];
            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_2289_cast_fp16_0, y = var_2302_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2305_to_fp16 = const()[name = string("op_2305_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_2305_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_2309 = const()[name = string("op_2309"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_87_cast_fp16 = softmax(axis = var_2309, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2315_transpose_x_1 = const()[name = string("op_2315_transpose_x_1"), val = bool(true)];
            bool var_2315_transpose_y_1 = const()[name = string("op_2315_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2315_cast_fp16 = matmul(transpose_x = var_2315_transpose_x_1, transpose_y = var_2315_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2299_cast_fp16_0)[name = string("op_2315_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_2289_cast_fp16_1, y = var_2302_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2317_to_fp16 = const()[name = string("op_2317_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_2317_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_2321 = const()[name = string("op_2321"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_95_cast_fp16 = softmax(axis = var_2321, 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_2299_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2329 = const()[name = string("op_2329"), 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_2329, interleave = attn_output_43_interleave_0, values = (var_2315_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2333_perm_0 = const()[name = string("op_2333_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_2333_cast_fp16 = transpose(perm = var_2333_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_47_cast_fp16 = reshape(shape = concat_71x, x = var_2333_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_2366_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_2366_cast_fp16")];
            int32 var_2364 = const()[name = string("op_2364"), 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_2364, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2366_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(758742592)))];
            fp16 var_2376_to_fp16 = const()[name = string("op_2376_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_2376_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2387_split_sizes_0 = const()[name = string("op_2387_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2387_axis_0 = const()[name = string("op_2387_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2387_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2387_cast_fp16_1 = split(axis = var_2387_axis_0, split_sizes = var_2387_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2387_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(758750848)))];
            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_to_fp16, x = var_2387_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2404_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [2]> var_2410_strides_0 = const()[name = string("op_2410_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2410_pad_type_0 = const()[name = string("op_2410_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2410_pad_0 = const()[name = string("op_2410_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2410_dilations_0 = const()[name = string("op_2410_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2410_groups_0 = const()[name = string("op_2410_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2410_cast_fp16 = conv(dilations = var_2410_dilations_0, groups = var_2410_groups_0, pad = var_2410_pad_0, pad_type = var_2410_pad_type_0, strides = var_2410_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2387_cast_fp16_0)[name = string("op_2410_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_59_cast_fp16 = mul(x = var_2404_cast_fp16, y = var_2410_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_2428_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_2428_cast_fp16")];
            int32 var_2426 = const()[name = string("op_2426"), 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_2426, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2428_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(783916736)))];
            fp16 var_2438_to_fp16 = const()[name = string("op_2438_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_2438_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2449_split_sizes_0 = const()[name = string("op_2449_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2449_axis_0 = const()[name = string("op_2449_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2449_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2449_cast_fp16_1 = split(axis = var_2449_axis_0, split_sizes = var_2449_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2449_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(783924992)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792313664)))];
            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_to_fp16, x = var_2449_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            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_cast_fp16, x = var_2449_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_2506_cast_fp16 = reshape(shape = concat_73x, x = key_states_61_cast_fp16)[name = string("op_2506_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_2513_cast_fp16 = reshape(shape = concat_74x, x = value_states_37_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2517_cast_fp16 = mul(x = x_61_cast_fp16, y = var_452_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<int32, [2]> var_2518_split_sizes_0 = const()[name = string("op_2518_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2518_axis_0 = const()[name = string("op_2518_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2518_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2518_cast_fp16_1 = split(axis = var_2518_axis_0, split_sizes = var_2518_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2518_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_2520_cast_fp16 = mul(x = var_2518_cast_fp16_1, y = const_64_promoted_to_fp16)[name = string("op_2520_cast_fp16")];
            int32 var_2522 = const()[name = string("op_2522"), val = int32(-2)];
            bool var_2523_interleave_0 = const()[name = string("op_2523_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2523_cast_fp16 = concat(axis = var_2522, interleave = var_2523_interleave_0, values = (var_2520_cast_fp16, var_2518_cast_fp16_0))[name = string("op_2523_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2524_cast_fp16 = mul(x = var_2523_cast_fp16, y = var_459_cast_fp16)[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_39_cast_fp16 = add(x = var_2517_cast_fp16, y = var_2524_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2530_cast_fp16 = mul(x = var_2506_cast_fp16, y = var_452_cast_fp16)[name = string("op_2530_cast_fp16")];
            tensor<int32, [2]> var_2531_split_sizes_0 = const()[name = string("op_2531_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2531_axis_0 = const()[name = string("op_2531_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2531_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2531_cast_fp16_1 = split(axis = var_2531_axis_0, split_sizes = var_2531_split_sizes_0, x = var_2506_cast_fp16)[name = string("op_2531_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_2533_cast_fp16 = mul(x = var_2531_cast_fp16_1, y = const_65_promoted_to_fp16)[name = string("op_2533_cast_fp16")];
            int32 var_2535 = const()[name = string("op_2535"), val = int32(-2)];
            bool var_2536_interleave_0 = const()[name = string("op_2536_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2536_cast_fp16 = concat(axis = var_2535, interleave = var_2536_interleave_0, values = (var_2533_cast_fp16, var_2531_cast_fp16_0))[name = string("op_2536_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2537_cast_fp16 = mul(x = var_2536_cast_fp16, y = var_459_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_65_cast_fp16 = add(x = var_2530_cast_fp16, y = var_2537_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_23")];
            tensor<fp16, [14, 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, [14, 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_2513_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [14, 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, [14, 2, 2048, 128]> coreml_update_state_13 = read_state(input = value_cache)[name = string("coreml_update_state_13")];
            tensor<int32, [4]> var_2607_begin_0 = const()[name = string("op_2607_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2607_end_0 = const()[name = string("op_2607_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2607_end_mask_0 = const()[name = string("op_2607_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2607_cast_fp16 = slice_by_index(begin = var_2607_begin_0, end = var_2607_end_0, end_mask = var_2607_end_mask_0, x = coreml_update_state_12)[name = string("op_2607_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2610_axis_0 = const()[name = string("op_2610_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2610_cast_fp16_1 = split(axis = var_2610_axis_0, split_sizes = tile_12, x = var_2607_cast_fp16)[name = string("op_2610_cast_fp16")];
            tensor<int32, [4]> var_2617_begin_0 = const()[name = string("op_2617_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2617_end_0 = const()[name = string("op_2617_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_2617_end_mask_0 = const()[name = string("op_2617_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2617_cast_fp16 = slice_by_index(begin = var_2617_begin_0, end = var_2617_end_0, end_mask = var_2617_end_mask_0, x = coreml_update_state_13)[name = string("op_2617_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2620_axis_0 = const()[name = string("op_2620_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2620_cast_fp16_1 = split(axis = var_2620_axis_0, split_sizes = tile_13, x = var_2617_cast_fp16)[name = string("op_2620_cast_fp16")];
            tensor<int32, [2]> var_2623_split_sizes_0 = const()[name = string("op_2623_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2623_axis_0 = const()[name = string("op_2623_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2623_0, tensor<fp16, [1, 8, 128, 1]> var_2623_1 = split(axis = var_2623_axis_0, split_sizes = var_2623_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_2623")];
            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_2610_cast_fp16_0, y = var_2623_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_2626_to_fp16 = const()[name = string("op_2626_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_2626_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_2630 = const()[name = string("op_2630"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_103_cast_fp16 = softmax(axis = var_2630, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_2636_transpose_x_1 = const()[name = string("op_2636_transpose_x_1"), val = bool(true)];
            bool var_2636_transpose_y_1 = const()[name = string("op_2636_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2636_cast_fp16 = matmul(transpose_x = var_2636_transpose_x_1, transpose_y = var_2636_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_2620_cast_fp16_0)[name = string("op_2636_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_2610_cast_fp16_1, y = var_2623_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_2638_to_fp16 = const()[name = string("op_2638_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_2638_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_2642 = const()[name = string("op_2642"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_111_cast_fp16 = softmax(axis = var_2642, 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_2620_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_2650 = const()[name = string("op_2650"), 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_2650, interleave = attn_output_51_interleave_0, values = (var_2636_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_2654_perm_0 = const()[name = string("op_2654_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_2654_cast_fp16 = transpose(perm = var_2654_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_55_cast_fp16 = reshape(shape = concat_83x, x = var_2654_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_2687_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2687_cast_fp16")];
            int32 var_2685 = const()[name = string("op_2685"), 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_2685, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_2687_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(793362304)))];
            fp16 var_2697_to_fp16 = const()[name = string("op_2697_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_2697_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_2708_split_sizes_0 = const()[name = string("op_2708_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2708_axis_0 = const()[name = string("op_2708_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2708_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2708_cast_fp16_1 = split(axis = var_2708_axis_0, split_sizes = var_2708_split_sizes_0, x = out_27_cast_fp16)[name = string("op_2708_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_2708_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_2725_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_2725_cast_fp16")];
            tensor<int32, [2]> var_2731_strides_0 = const()[name = string("op_2731_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_2731_pad_type_0 = const()[name = string("op_2731_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_2731_pad_0 = const()[name = string("op_2731_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2731_dilations_0 = const()[name = string("op_2731_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2731_groups_0 = const()[name = string("op_2731_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_2731_cast_fp16 = conv(dilations = var_2731_dilations_0, groups = var_2731_groups_0, pad = var_2731_pad_0, pad_type = var_2731_pad_type_0, strides = var_2731_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_2708_cast_fp16_0)[name = string("op_2731_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_69_cast_fp16 = mul(x = var_2725_cast_fp16, y = var_2731_cast_fp16)[name = string("x_69_cast_fp16")];
            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_cast_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_2749_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_2749_cast_fp16")];
            int32 var_2747 = const()[name = string("op_2747"), 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_2747, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_2749_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(793370560)))];
            fp16 var_2759_to_fp16 = const()[name = string("op_2759_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_2759_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_2770_split_sizes_0 = const()[name = string("op_2770_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2770_axis_0 = const()[name = string("op_2770_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_2770_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_2770_cast_fp16_1 = split(axis = var_2770_axis_0, split_sizes = var_2770_split_sizes_0, x = out_29_cast_fp16)[name = string("op_2770_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(793378816)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(801767488)))];
            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_to_fp16, x = var_2770_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            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_cast_fp16, x = var_2770_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_2827_cast_fp16 = reshape(shape = concat_85x, x = key_states_71_cast_fp16)[name = string("op_2827_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_2834_cast_fp16 = reshape(shape = concat_86x, x = value_states_43_cast_fp16)[name = string("op_2834_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2838_cast_fp16 = mul(x = x_71_cast_fp16, y = var_452_cast_fp16)[name = string("op_2838_cast_fp16")];
            tensor<int32, [2]> var_2839_split_sizes_0 = const()[name = string("op_2839_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2839_axis_0 = const()[name = string("op_2839_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_2839_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_2839_cast_fp16_1 = split(axis = var_2839_axis_0, split_sizes = var_2839_split_sizes_0, x = x_71_cast_fp16)[name = string("op_2839_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_2841_cast_fp16 = mul(x = var_2839_cast_fp16_1, y = const_74_promoted_to_fp16)[name = string("op_2841_cast_fp16")];
            int32 var_2843 = const()[name = string("op_2843"), val = int32(-2)];
            bool var_2844_interleave_0 = const()[name = string("op_2844_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_2844_cast_fp16 = concat(axis = var_2843, interleave = var_2844_interleave_0, values = (var_2841_cast_fp16, var_2839_cast_fp16_0))[name = string("op_2844_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_2845_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_459_cast_fp16)[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_45_cast_fp16 = add(x = var_2838_cast_fp16, y = var_2845_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2851_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_452_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [2]> var_2852_split_sizes_0 = const()[name = string("op_2852_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2852_axis_0 = const()[name = string("op_2852_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_2852_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_2852_cast_fp16_1 = split(axis = var_2852_axis_0, split_sizes = var_2852_split_sizes_0, x = var_2827_cast_fp16)[name = string("op_2852_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_2854_cast_fp16 = mul(x = var_2852_cast_fp16_1, y = const_75_promoted_to_fp16)[name = string("op_2854_cast_fp16")];
            int32 var_2856 = const()[name = string("op_2856"), val = int32(-2)];
            bool var_2857_interleave_0 = const()[name = string("op_2857_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_2857_cast_fp16 = concat(axis = var_2856, interleave = var_2857_interleave_0, values = (var_2854_cast_fp16, var_2852_cast_fp16_0))[name = string("op_2857_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2858_cast_fp16 = mul(x = var_2857_cast_fp16, y = var_459_cast_fp16)[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_75_cast_fp16 = add(x = var_2851_cast_fp16, y = var_2858_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_20")];
            tensor<fp16, [14, 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, [14, 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_2834_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [14, 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, [14, 2, 2048, 128]> coreml_update_state_15 = read_state(input = value_cache)[name = string("coreml_update_state_15")];
            tensor<int32, [4]> var_2928_begin_0 = const()[name = string("op_2928_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2928_end_0 = const()[name = string("op_2928_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2928_end_mask_0 = const()[name = string("op_2928_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2928_cast_fp16 = slice_by_index(begin = var_2928_begin_0, end = var_2928_end_0, end_mask = var_2928_end_mask_0, x = coreml_update_state_14)[name = string("op_2928_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2931_axis_0 = const()[name = string("op_2931_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2931_cast_fp16_1 = split(axis = var_2931_axis_0, split_sizes = tile_14, x = var_2928_cast_fp16)[name = string("op_2931_cast_fp16")];
            tensor<int32, [4]> var_2938_begin_0 = const()[name = string("op_2938_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_2938_end_0 = const()[name = string("op_2938_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_2938_end_mask_0 = const()[name = string("op_2938_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2938_cast_fp16 = slice_by_index(begin = var_2938_begin_0, end = var_2938_end_0, end_mask = var_2938_end_mask_0, x = coreml_update_state_15)[name = string("op_2938_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2941_axis_0 = const()[name = string("op_2941_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2941_cast_fp16_1 = split(axis = var_2941_axis_0, split_sizes = tile_15, x = var_2938_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<int32, [2]> var_2944_split_sizes_0 = const()[name = string("op_2944_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2944_axis_0 = const()[name = string("op_2944_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_2944_0, tensor<fp16, [1, 8, 128, 1]> var_2944_1 = split(axis = var_2944_axis_0, split_sizes = var_2944_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_2944")];
            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_2931_cast_fp16_0, y = var_2944_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_2947_to_fp16 = const()[name = string("op_2947_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_2947_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_2951 = const()[name = string("op_2951"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_119_cast_fp16 = softmax(axis = var_2951, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_2957_transpose_x_1 = const()[name = string("op_2957_transpose_x_1"), val = bool(true)];
            bool var_2957_transpose_y_1 = const()[name = string("op_2957_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_2957_cast_fp16 = matmul(transpose_x = var_2957_transpose_x_1, transpose_y = var_2957_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_2941_cast_fp16_0)[name = string("op_2957_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_2931_cast_fp16_1, y = var_2944_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_2959_to_fp16 = const()[name = string("op_2959_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_2959_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_2963 = const()[name = string("op_2963"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_127_cast_fp16 = softmax(axis = var_2963, 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_2941_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_2971 = const()[name = string("op_2971"), 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_2971, interleave = attn_output_59_interleave_0, values = (var_2957_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_2975_perm_0 = const()[name = string("op_2975_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_2975_cast_fp16 = transpose(perm = var_2975_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_63_cast_fp16 = reshape(shape = concat_95x, x = var_2975_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_3008_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_3008_cast_fp16")];
            int32 var_3006 = const()[name = string("op_3006"), 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_3006, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3008_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(802816128)))];
            fp16 var_3018_to_fp16 = const()[name = string("op_3018_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_3018_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3029_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3029_cast_fp16_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3029_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_3029_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3046_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3046_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(802824384)))];
            tensor<int32, [2]> var_3052_strides_0 = const()[name = string("op_3052_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3052_pad_type_0 = const()[name = string("op_3052_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3052_pad_0 = const()[name = string("op_3052_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3052_dilations_0 = const()[name = string("op_3052_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3052_groups_0 = const()[name = string("op_3052_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3052_cast_fp16 = conv(dilations = var_3052_dilations_0, groups = var_3052_groups_0, pad = var_3052_pad_0, pad_type = var_3052_pad_type_0, strides = var_3052_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_3029_cast_fp16_0)[name = string("op_3052_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_79_cast_fp16 = mul(x = var_3046_cast_fp16, y = var_3052_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(827990272)))];
            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_3070_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_3070_cast_fp16")];
            int32 var_3068 = const()[name = string("op_3068"), 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_3068, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3070_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(853156160)))];
            fp16 var_3080_to_fp16 = const()[name = string("op_3080_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_3080_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3091_split_sizes_0 = const()[name = string("op_3091_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3091_axis_0 = const()[name = string("op_3091_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3091_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3091_cast_fp16_1 = split(axis = var_3091_axis_0, split_sizes = var_3091_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3091_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(853164416)))];
            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_3091_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(861553088)))];
            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_3091_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            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_cast_fp16, x = var_3091_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_3148_cast_fp16 = reshape(shape = concat_97x, x = key_states_81_cast_fp16)[name = string("op_3148_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_3155_cast_fp16 = reshape(shape = concat_98x, x = value_states_49_cast_fp16)[name = string("op_3155_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3159_cast_fp16 = mul(x = x_81_cast_fp16, y = var_452_cast_fp16)[name = string("op_3159_cast_fp16")];
            tensor<int32, [2]> var_3160_split_sizes_0 = const()[name = string("op_3160_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3160_axis_0 = const()[name = string("op_3160_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3160_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3160_cast_fp16_1 = split(axis = var_3160_axis_0, split_sizes = var_3160_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3160_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_3162_cast_fp16 = mul(x = var_3160_cast_fp16_1, y = const_84_promoted_to_fp16)[name = string("op_3162_cast_fp16")];
            int32 var_3164 = const()[name = string("op_3164"), val = int32(-2)];
            bool var_3165_interleave_0 = const()[name = string("op_3165_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3165_cast_fp16 = concat(axis = var_3164, interleave = var_3165_interleave_0, values = (var_3162_cast_fp16, var_3160_cast_fp16_0))[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3166_cast_fp16 = mul(x = var_3165_cast_fp16, y = var_459_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_51_cast_fp16 = add(x = var_3159_cast_fp16, y = var_3166_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3172_cast_fp16 = mul(x = var_3148_cast_fp16, y = var_452_cast_fp16)[name = string("op_3172_cast_fp16")];
            tensor<int32, [2]> var_3173_split_sizes_0 = const()[name = string("op_3173_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3173_axis_0 = const()[name = string("op_3173_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3173_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3173_cast_fp16_1 = split(axis = var_3173_axis_0, split_sizes = var_3173_split_sizes_0, x = var_3148_cast_fp16)[name = string("op_3173_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_3175_cast_fp16 = mul(x = var_3173_cast_fp16_1, y = const_85_promoted_to_fp16)[name = string("op_3175_cast_fp16")];
            int32 var_3177 = const()[name = string("op_3177"), val = int32(-2)];
            bool var_3178_interleave_0 = const()[name = string("op_3178_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3178_cast_fp16 = concat(axis = var_3177, interleave = var_3178_interleave_0, values = (var_3175_cast_fp16, var_3173_cast_fp16_0))[name = string("op_3178_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3179_cast_fp16 = mul(x = var_3178_cast_fp16, y = var_459_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_85_cast_fp16 = add(x = var_3172_cast_fp16, y = var_3179_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_17")];
            tensor<fp16, [14, 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, [14, 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_3155_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [14, 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, [14, 2, 2048, 128]> coreml_update_state_17 = read_state(input = value_cache)[name = string("coreml_update_state_17")];
            tensor<int32, [4]> var_3249_begin_0 = const()[name = string("op_3249_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3249_end_0 = const()[name = string("op_3249_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3249_end_mask_0 = const()[name = string("op_3249_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3249_cast_fp16 = slice_by_index(begin = var_3249_begin_0, end = var_3249_end_0, end_mask = var_3249_end_mask_0, x = coreml_update_state_16)[name = string("op_3249_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3252_axis_0 = const()[name = string("op_3252_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3252_cast_fp16_1 = split(axis = var_3252_axis_0, split_sizes = tile_16, x = var_3249_cast_fp16)[name = string("op_3252_cast_fp16")];
            tensor<int32, [4]> var_3259_begin_0 = const()[name = string("op_3259_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3259_end_0 = const()[name = string("op_3259_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3259_end_mask_0 = const()[name = string("op_3259_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3259_cast_fp16 = slice_by_index(begin = var_3259_begin_0, end = var_3259_end_0, end_mask = var_3259_end_mask_0, x = coreml_update_state_17)[name = string("op_3259_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3262_axis_0 = const()[name = string("op_3262_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3262_cast_fp16_1 = split(axis = var_3262_axis_0, split_sizes = tile_17, x = var_3259_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<int32, [2]> var_3265_split_sizes_0 = const()[name = string("op_3265_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3265_axis_0 = const()[name = string("op_3265_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3265_0, tensor<fp16, [1, 8, 128, 1]> var_3265_1 = split(axis = var_3265_axis_0, split_sizes = var_3265_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3265")];
            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_3252_cast_fp16_0, y = var_3265_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3268_to_fp16 = const()[name = string("op_3268_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_3268_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_3272 = const()[name = string("op_3272"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_135_cast_fp16 = softmax(axis = var_3272, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3278_transpose_x_1 = const()[name = string("op_3278_transpose_x_1"), val = bool(true)];
            bool var_3278_transpose_y_1 = const()[name = string("op_3278_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3278_cast_fp16 = matmul(transpose_x = var_3278_transpose_x_1, transpose_y = var_3278_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3262_cast_fp16_0)[name = string("op_3278_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_3252_cast_fp16_1, y = var_3265_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3280_to_fp16 = const()[name = string("op_3280_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_3280_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_3284 = const()[name = string("op_3284"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_143_cast_fp16 = softmax(axis = var_3284, x = attn_weights_141_cast_fp16)[name = string("attn_weights_143_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_143_cast_fp16, y = var_3262_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3292 = const()[name = string("op_3292"), 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_3292, interleave = attn_output_67_interleave_0, values = (var_3278_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3296_perm_0 = const()[name = string("op_3296_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_3296_cast_fp16 = transpose(perm = var_3296_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_71_cast_fp16 = reshape(shape = concat_107x, x = var_3296_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            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_cast_fp16, x = attn_output_71_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_3329_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_3329_cast_fp16")];
            int32 var_3327 = const()[name = string("op_3327"), 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_3327, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3329_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(862601728)))];
            fp16 var_3339_to_fp16 = const()[name = string("op_3339_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_3339_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3350_split_sizes_0 = const()[name = string("op_3350_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3350_axis_0 = const()[name = string("op_3350_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3350_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3350_cast_fp16_1 = split(axis = var_3350_axis_0, split_sizes = var_3350_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3350_cast_fp16")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_8_mlp_gate_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3367_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3367_cast_fp16")];
            tensor<int32, [2]> var_3373_strides_0 = const()[name = string("op_3373_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3373_pad_type_0 = const()[name = string("op_3373_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3373_pad_0 = const()[name = string("op_3373_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3373_dilations_0 = const()[name = string("op_3373_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3373_groups_0 = const()[name = string("op_3373_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3373_cast_fp16 = conv(dilations = var_3373_dilations_0, groups = var_3373_groups_0, pad = var_3373_pad_0, pad_type = var_3373_pad_type_0, strides = var_3373_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3350_cast_fp16_0)[name = string("op_3373_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_89_cast_fp16 = mul(x = var_3367_cast_fp16, y = var_3373_cast_fp16)[name = string("x_89_cast_fp16")];
            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_cast_fp16, x = x_89_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_89_cast_fp16 = add(x = hidden_states_85_cast_fp16, y = hidden_states_87_cast_fp16)[name = string("hidden_states_89_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_3391_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_3391_cast_fp16")];
            int32 var_3389 = const()[name = string("op_3389"), 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_3389, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3391_cast_fp16))[name = string("doubled_73_cast_fp16")];
            tensor<int32, [1]> out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_37_gamma_0_to_fp16 = const()[name = string("out_37_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862609984)))];
            fp16 var_3401_to_fp16 = const()[name = string("op_3401_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_3401_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3412_split_sizes_0 = const()[name = string("op_3412_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3412_axis_0 = const()[name = string("op_3412_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3412_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3412_cast_fp16_1 = split(axis = var_3412_axis_0, split_sizes = var_3412_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3412_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(862618240)))];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_55_pad_0 = const()[name = string("query_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_55_dilations_0 = const()[name = string("query_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_55_groups_0 = const()[name = string("query_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_55_cast_fp16 = conv(dilations = query_states_55_dilations_0, groups = query_states_55_groups_0, pad = query_states_55_pad_0, pad_type = query_states_55_pad_type_0, strides = query_states_55_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(871006912)))];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_91_pad_0 = const()[name = string("key_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_91_dilations_0 = const()[name = string("key_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_91_groups_0 = const()[name = string("key_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_91_cast_fp16 = conv(dilations = key_states_91_dilations_0, groups = key_states_91_groups_0, pad = key_states_91_pad_0, pad_type = key_states_91_pad_type_0, strides = key_states_91_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = var_3412_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_55_pad_0 = const()[name = string("value_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_55_dilations_0 = const()[name = string("value_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_55_groups_0 = const()[name = string("value_states_55_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_55_cast_fp16 = conv(dilations = value_states_55_dilations_0, groups = value_states_55_groups_0, pad = value_states_55_pad_0, pad_type = value_states_55_pad_type_0, strides = value_states_55_strides_0, weight = layers_9_self_attn_v_proj_weight_cast_fp16, x = var_3412_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_91_cast_fp16 = reshape(shape = concat_108x, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3469_cast_fp16 = reshape(shape = concat_109x, x = key_states_91_cast_fp16)[name = string("op_3469_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3476_cast_fp16 = reshape(shape = concat_110x, x = value_states_55_cast_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3480_cast_fp16 = mul(x = x_91_cast_fp16, y = var_452_cast_fp16)[name = string("op_3480_cast_fp16")];
            tensor<int32, [2]> var_3481_split_sizes_0 = const()[name = string("op_3481_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3481_axis_0 = const()[name = string("op_3481_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3481_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3481_cast_fp16_1 = split(axis = var_3481_axis_0, split_sizes = var_3481_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3481_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_3483_cast_fp16 = mul(x = var_3481_cast_fp16_1, y = const_94_promoted_to_fp16)[name = string("op_3483_cast_fp16")];
            int32 var_3485 = const()[name = string("op_3485"), val = int32(-2)];
            bool var_3486_interleave_0 = const()[name = string("op_3486_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3486_cast_fp16 = concat(axis = var_3485, interleave = var_3486_interleave_0, values = (var_3483_cast_fp16, var_3481_cast_fp16_0))[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3487_cast_fp16 = mul(x = var_3486_cast_fp16, y = var_459_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_57_cast_fp16 = add(x = var_3480_cast_fp16, y = var_3487_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3493_cast_fp16 = mul(x = var_3469_cast_fp16, y = var_452_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [2]> var_3494_split_sizes_0 = const()[name = string("op_3494_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3494_axis_0 = const()[name = string("op_3494_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3494_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3494_cast_fp16_1 = split(axis = var_3494_axis_0, split_sizes = var_3494_split_sizes_0, x = var_3469_cast_fp16)[name = string("op_3494_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_3496_cast_fp16 = mul(x = var_3494_cast_fp16_1, y = const_95_promoted_to_fp16)[name = string("op_3496_cast_fp16")];
            int32 var_3498 = const()[name = string("op_3498"), val = int32(-2)];
            bool var_3499_interleave_0 = const()[name = string("op_3499_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3499_cast_fp16 = concat(axis = var_3498, interleave = var_3499_interleave_0, values = (var_3496_cast_fp16, var_3494_cast_fp16_0))[name = string("op_3499_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3500_cast_fp16 = mul(x = var_3499_cast_fp16, y = var_459_cast_fp16)[name = string("op_3500_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_95_cast_fp16 = add(x = var_3493_cast_fp16, y = var_3500_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (expand_dims_108, expand_dims_109, position_id, expand_dims_111))[name = string("concat_113")];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> concat_114_values1_0 = const()[name = string("concat_114_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_114_values3_0 = const()[name = string("concat_114_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_114_axis_0 = const()[name = string("concat_114_axis_0"), val = int32(0)];
            bool concat_114_interleave_0 = const()[name = string("concat_114_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_114 = concat(axis = concat_114_axis_0, interleave = concat_114_interleave_0, values = (expand_dims_112, concat_114_values1_0, cache_position_end, concat_114_values3_0))[name = string("concat_114")];
            tensor<int32, [4]> key_states_97_perm_0 = const()[name = string("key_states_97_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_97_cast_fp16 = transpose(perm = key_states_97_perm_0, x = key_states_95_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = key_states_97_cast_fp16, x = coreml_update_state_16)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_18_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_18 = read_state(input = key_cache)[name = string("coreml_update_state_18")];
            tensor<int32, [4]> value_states_57_perm_0 = const()[name = string("value_states_57_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_57_cast_fp16 = transpose(perm = value_states_57_perm_0, x = var_3476_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_113, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_114, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = value_states_57_cast_fp16, x = coreml_update_state_17)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_19_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_19 = read_state(input = value_cache)[name = string("coreml_update_state_19")];
            tensor<int32, [4]> var_3570_begin_0 = const()[name = string("op_3570_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3570_end_0 = const()[name = string("op_3570_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3570_end_mask_0 = const()[name = string("op_3570_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = coreml_update_state_18)[name = string("op_3570_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3573_axis_0 = const()[name = string("op_3573_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3573_cast_fp16_1 = split(axis = var_3573_axis_0, split_sizes = tile_18, x = var_3570_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [4]> var_3580_begin_0 = const()[name = string("op_3580_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3580_end_0 = const()[name = string("op_3580_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3580_end_mask_0 = const()[name = string("op_3580_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3580_cast_fp16 = slice_by_index(begin = var_3580_begin_0, end = var_3580_end_0, end_mask = var_3580_end_mask_0, x = coreml_update_state_19)[name = string("op_3580_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3583_axis_0 = const()[name = string("op_3583_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3583_cast_fp16_1 = split(axis = var_3583_axis_0, split_sizes = tile_19, x = var_3580_cast_fp16)[name = string("op_3583_cast_fp16")];
            tensor<int32, [2]> var_3586_split_sizes_0 = const()[name = string("op_3586_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3586_axis_0 = const()[name = string("op_3586_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3586_0, tensor<fp16, [1, 8, 128, 1]> var_3586_1 = split(axis = var_3586_axis_0, split_sizes = var_3586_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3586")];
            bool attn_weights_145_transpose_x_0 = const()[name = string("attn_weights_145_transpose_x_0"), val = bool(false)];
            bool attn_weights_145_transpose_y_0 = const()[name = string("attn_weights_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_145_cast_fp16 = matmul(transpose_x = attn_weights_145_transpose_x_0, transpose_y = attn_weights_145_transpose_y_0, x = var_3573_cast_fp16_0, y = var_3586_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3589_to_fp16 = const()[name = string("op_3589_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_147_cast_fp16 = mul(x = attn_weights_145_cast_fp16, y = var_3589_to_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_149_cast_fp16 = add(x = attn_weights_147_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_149_cast_fp16")];
            int32 var_3593 = const()[name = string("op_3593"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_151_cast_fp16 = softmax(axis = var_3593, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_3599_transpose_x_1 = const()[name = string("op_3599_transpose_x_1"), val = bool(true)];
            bool var_3599_transpose_y_1 = const()[name = string("op_3599_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3599_cast_fp16 = matmul(transpose_x = var_3599_transpose_x_1, transpose_y = var_3599_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3583_cast_fp16_0)[name = string("op_3599_cast_fp16")];
            bool attn_weights_153_transpose_x_0 = const()[name = string("attn_weights_153_transpose_x_0"), val = bool(false)];
            bool attn_weights_153_transpose_y_0 = const()[name = string("attn_weights_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_153_cast_fp16 = matmul(transpose_x = attn_weights_153_transpose_x_0, transpose_y = attn_weights_153_transpose_y_0, x = var_3573_cast_fp16_1, y = var_3586_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_3601_to_fp16 = const()[name = string("op_3601_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_155_cast_fp16 = mul(x = attn_weights_153_cast_fp16, y = var_3601_to_fp16)[name = string("attn_weights_155_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_157_cast_fp16 = add(x = attn_weights_155_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_157_cast_fp16")];
            int32 var_3605 = const()[name = string("op_3605"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_159_cast_fp16 = softmax(axis = var_3605, x = attn_weights_157_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            bool attn_output_73_transpose_x_1 = const()[name = string("attn_output_73_transpose_x_1"), val = bool(true)];
            bool attn_output_73_transpose_y_1 = const()[name = string("attn_output_73_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_1, transpose_y = attn_output_73_transpose_y_1, x = attn_weights_159_cast_fp16, y = var_3583_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_3613 = const()[name = string("op_3613"), val = int32(1)];
            bool attn_output_75_interleave_0 = const()[name = string("attn_output_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_75_cast_fp16 = concat(axis = var_3613, interleave = attn_output_75_interleave_0, values = (var_3599_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_3617_perm_0 = const()[name = string("op_3617_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_3617_cast_fp16 = transpose(perm = var_3617_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_79_cast_fp16 = reshape(shape = concat_119x, x = var_3617_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_93_pad_0 = const()[name = string("hidden_states_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_93_dilations_0 = const()[name = string("hidden_states_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_93_groups_0 = const()[name = string("hidden_states_93_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_93_cast_fp16 = conv(dilations = hidden_states_93_dilations_0, groups = hidden_states_93_groups_0, pad = hidden_states_93_pad_0, pad_type = hidden_states_93_pad_type_0, strides = hidden_states_93_strides_0, weight = layers_9_self_attn_o_proj_weight_cast_fp16, x = attn_output_79_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_95_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3650_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_3650_cast_fp16")];
            int32 var_3648 = const()[name = string("op_3648"), val = int32(1)];
            bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_77_cast_fp16 = concat(axis = var_3648, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_3650_cast_fp16))[name = string("doubled_77_cast_fp16")];
            tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872055552)))];
            fp16 var_3660_to_fp16 = const()[name = string("op_3660_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_3660_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_3671_split_sizes_0 = const()[name = string("op_3671_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3671_axis_0 = const()[name = string("op_3671_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3671_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3671_cast_fp16_1 = split(axis = var_3671_axis_0, split_sizes = var_3671_split_sizes_0, x = out_39_cast_fp16)[name = string("op_3671_cast_fp16")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_9_mlp_gate_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_3688_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_3688_cast_fp16")];
            tensor<int32, [2]> var_3694_strides_0 = const()[name = string("op_3694_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_3694_pad_type_0 = const()[name = string("op_3694_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_3694_pad_0 = const()[name = string("op_3694_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3694_dilations_0 = const()[name = string("op_3694_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3694_groups_0 = const()[name = string("op_3694_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_3694_cast_fp16 = conv(dilations = var_3694_dilations_0, groups = var_3694_groups_0, pad = var_3694_pad_0, pad_type = var_3694_pad_type_0, strides = var_3694_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_3671_cast_fp16_0)[name = string("op_3694_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_99_cast_fp16 = mul(x = var_3688_cast_fp16, y = var_3694_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_97_pad_0 = const()[name = string("hidden_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_97_dilations_0 = const()[name = string("hidden_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_97_groups_0 = const()[name = string("hidden_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_97_cast_fp16 = conv(dilations = hidden_states_97_dilations_0, groups = hidden_states_97_groups_0, pad = hidden_states_97_pad_0, pad_type = hidden_states_97_pad_type_0, strides = hidden_states_97_strides_0, weight = layers_9_mlp_down_proj_weight_cast_fp16, x = x_99_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_99_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = hidden_states_97_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3712_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_3712_cast_fp16")];
            int32 var_3710 = const()[name = string("op_3710"), val = int32(1)];
            bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_81_cast_fp16 = concat(axis = var_3710, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_3712_cast_fp16))[name = string("doubled_81_cast_fp16")];
            tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872063808)))];
            fp16 var_3722_to_fp16 = const()[name = string("op_3722_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_3722_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_3733_split_sizes_0 = const()[name = string("op_3733_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3733_axis_0 = const()[name = string("op_3733_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3733_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3733_cast_fp16_1 = split(axis = var_3733_axis_0, split_sizes = var_3733_split_sizes_0, x = out_41_cast_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872072064)))];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_61_pad_0 = const()[name = string("query_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_61_dilations_0 = const()[name = string("query_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_61_groups_0 = const()[name = string("query_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_61_cast_fp16 = conv(dilations = query_states_61_dilations_0, groups = query_states_61_groups_0, pad = query_states_61_pad_0, pad_type = query_states_61_pad_type_0, strides = query_states_61_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(880460736)))];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_101_pad_0 = const()[name = string("key_states_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_101_dilations_0 = const()[name = string("key_states_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_101_groups_0 = const()[name = string("key_states_101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_101_cast_fp16 = conv(dilations = key_states_101_dilations_0, groups = key_states_101_groups_0, pad = key_states_101_pad_0, pad_type = key_states_101_pad_type_0, strides = key_states_101_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = var_3733_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_61_pad_0 = const()[name = string("value_states_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_61_dilations_0 = const()[name = string("value_states_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_61_groups_0 = const()[name = string("value_states_61_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_61_cast_fp16 = conv(dilations = value_states_61_dilations_0, groups = value_states_61_groups_0, pad = value_states_61_pad_0, pad_type = value_states_61_pad_type_0, strides = value_states_61_strides_0, weight = layers_10_self_attn_v_proj_weight_cast_fp16, x = var_3733_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_101_cast_fp16 = reshape(shape = concat_120x, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3790_cast_fp16 = reshape(shape = concat_121x, x = key_states_101_cast_fp16)[name = string("op_3790_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_3797_cast_fp16 = reshape(shape = concat_122x, x = value_states_61_cast_fp16)[name = string("op_3797_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3801_cast_fp16 = mul(x = x_101_cast_fp16, y = var_452_cast_fp16)[name = string("op_3801_cast_fp16")];
            tensor<int32, [2]> var_3802_split_sizes_0 = const()[name = string("op_3802_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3802_axis_0 = const()[name = string("op_3802_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_3802_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_3802_cast_fp16_1 = split(axis = var_3802_axis_0, split_sizes = var_3802_split_sizes_0, x = x_101_cast_fp16)[name = string("op_3802_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_3804_cast_fp16 = mul(x = var_3802_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3804_cast_fp16")];
            int32 var_3806 = const()[name = string("op_3806"), val = int32(-2)];
            bool var_3807_interleave_0 = const()[name = string("op_3807_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_3807_cast_fp16 = concat(axis = var_3806, interleave = var_3807_interleave_0, values = (var_3804_cast_fp16, var_3802_cast_fp16_0))[name = string("op_3807_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_3808_cast_fp16 = mul(x = var_3807_cast_fp16, y = var_459_cast_fp16)[name = string("op_3808_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_63_cast_fp16 = add(x = var_3801_cast_fp16, y = var_3808_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3814_cast_fp16 = mul(x = var_3790_cast_fp16, y = var_452_cast_fp16)[name = string("op_3814_cast_fp16")];
            tensor<int32, [2]> var_3815_split_sizes_0 = const()[name = string("op_3815_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3815_axis_0 = const()[name = string("op_3815_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_3815_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_3815_cast_fp16_1 = split(axis = var_3815_axis_0, split_sizes = var_3815_split_sizes_0, x = var_3790_cast_fp16)[name = string("op_3815_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_3817_cast_fp16 = mul(x = var_3815_cast_fp16_1, y = const_105_promoted_to_fp16)[name = string("op_3817_cast_fp16")];
            int32 var_3819 = const()[name = string("op_3819"), val = int32(-2)];
            bool var_3820_interleave_0 = const()[name = string("op_3820_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_3820_cast_fp16 = concat(axis = var_3819, interleave = var_3820_interleave_0, values = (var_3817_cast_fp16, var_3815_cast_fp16_0))[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3821_cast_fp16 = mul(x = var_3820_cast_fp16, y = var_459_cast_fp16)[name = string("op_3821_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_105_cast_fp16 = add(x = var_3814_cast_fp16, y = var_3821_cast_fp16)[name = string("key_states_105_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (expand_dims_120, expand_dims_121, position_id, expand_dims_123))[name = string("concat_125")];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> concat_126_values1_0 = const()[name = string("concat_126_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_126_values3_0 = const()[name = string("concat_126_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_126_axis_0 = const()[name = string("concat_126_axis_0"), val = int32(0)];
            bool concat_126_interleave_0 = const()[name = string("concat_126_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_126 = concat(axis = concat_126_axis_0, interleave = concat_126_interleave_0, values = (expand_dims_124, concat_126_values1_0, cache_position_end, concat_126_values3_0))[name = string("concat_126")];
            tensor<int32, [4]> key_states_107_perm_0 = const()[name = string("key_states_107_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_107_cast_fp16 = transpose(perm = key_states_107_perm_0, x = key_states_105_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = key_states_107_cast_fp16, x = coreml_update_state_18)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_20_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_20 = read_state(input = key_cache)[name = string("coreml_update_state_20")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = string("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_3797_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_125, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_126, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = value_states_63_cast_fp16, x = coreml_update_state_19)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_21_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_21 = read_state(input = value_cache)[name = string("coreml_update_state_21")];
            tensor<int32, [4]> var_3891_begin_0 = const()[name = string("op_3891_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3891_end_0 = const()[name = string("op_3891_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3891_end_mask_0 = const()[name = string("op_3891_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3891_cast_fp16 = slice_by_index(begin = var_3891_begin_0, end = var_3891_end_0, end_mask = var_3891_end_mask_0, x = coreml_update_state_20)[name = string("op_3891_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3894_axis_0 = const()[name = string("op_3894_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3894_cast_fp16_1 = split(axis = var_3894_axis_0, split_sizes = tile_20, x = var_3891_cast_fp16)[name = string("op_3894_cast_fp16")];
            tensor<int32, [4]> var_3901_begin_0 = const()[name = string("op_3901_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_3901_end_0 = const()[name = string("op_3901_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_3901_end_mask_0 = const()[name = string("op_3901_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3901_cast_fp16 = slice_by_index(begin = var_3901_begin_0, end = var_3901_end_0, end_mask = var_3901_end_mask_0, x = coreml_update_state_21)[name = string("op_3901_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3904_axis_0 = const()[name = string("op_3904_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3904_cast_fp16_1 = split(axis = var_3904_axis_0, split_sizes = tile_21, x = var_3901_cast_fp16)[name = string("op_3904_cast_fp16")];
            tensor<int32, [2]> var_3907_split_sizes_0 = const()[name = string("op_3907_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3907_axis_0 = const()[name = string("op_3907_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_3907_0, tensor<fp16, [1, 8, 128, 1]> var_3907_1 = split(axis = var_3907_axis_0, split_sizes = var_3907_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_3907")];
            bool attn_weights_161_transpose_x_0 = const()[name = string("attn_weights_161_transpose_x_0"), val = bool(false)];
            bool attn_weights_161_transpose_y_0 = const()[name = string("attn_weights_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_161_cast_fp16 = matmul(transpose_x = attn_weights_161_transpose_x_0, transpose_y = attn_weights_161_transpose_y_0, x = var_3894_cast_fp16_0, y = var_3907_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_3910_to_fp16 = const()[name = string("op_3910_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_163_cast_fp16 = mul(x = attn_weights_161_cast_fp16, y = var_3910_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            int32 var_3914 = const()[name = string("op_3914"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_167_cast_fp16 = softmax(axis = var_3914, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_3920_transpose_x_1 = const()[name = string("op_3920_transpose_x_1"), val = bool(true)];
            bool var_3920_transpose_y_1 = const()[name = string("op_3920_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_3920_cast_fp16 = matmul(transpose_x = var_3920_transpose_x_1, transpose_y = var_3920_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_3904_cast_fp16_0)[name = string("op_3920_cast_fp16")];
            bool attn_weights_169_transpose_x_0 = const()[name = string("attn_weights_169_transpose_x_0"), val = bool(false)];
            bool attn_weights_169_transpose_y_0 = const()[name = string("attn_weights_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_169_cast_fp16 = matmul(transpose_x = attn_weights_169_transpose_x_0, transpose_y = attn_weights_169_transpose_y_0, x = var_3894_cast_fp16_1, y = var_3907_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_3922_to_fp16 = const()[name = string("op_3922_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_171_cast_fp16 = mul(x = attn_weights_169_cast_fp16, y = var_3922_to_fp16)[name = string("attn_weights_171_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_173_cast_fp16 = add(x = attn_weights_171_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_173_cast_fp16")];
            int32 var_3926 = const()[name = string("op_3926"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_175_cast_fp16 = softmax(axis = var_3926, x = attn_weights_173_cast_fp16)[name = string("attn_weights_175_cast_fp16")];
            bool attn_output_81_transpose_x_1 = const()[name = string("attn_output_81_transpose_x_1"), val = bool(true)];
            bool attn_output_81_transpose_y_1 = const()[name = string("attn_output_81_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_1, transpose_y = attn_output_81_transpose_y_1, x = attn_weights_175_cast_fp16, y = var_3904_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_3934 = const()[name = string("op_3934"), val = int32(1)];
            bool attn_output_83_interleave_0 = const()[name = string("attn_output_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_83_cast_fp16 = concat(axis = var_3934, interleave = attn_output_83_interleave_0, values = (var_3920_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_3938_perm_0 = const()[name = string("op_3938_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_131x = const()[name = string("concat_131x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_3938_cast_fp16 = transpose(perm = var_3938_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_87_cast_fp16 = reshape(shape = concat_131x, x = var_3938_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_103_pad_0 = const()[name = string("hidden_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_103_dilations_0 = const()[name = string("hidden_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_103_groups_0 = const()[name = string("hidden_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_103_cast_fp16 = conv(dilations = hidden_states_103_dilations_0, groups = hidden_states_103_groups_0, pad = hidden_states_103_pad_0, pad_type = hidden_states_103_pad_type_0, strides = hidden_states_103_strides_0, weight = layers_10_self_attn_o_proj_weight_cast_fp16, x = attn_output_87_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_105_cast_fp16 = add(x = hidden_states_99_cast_fp16, y = hidden_states_103_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_3971_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_3971_cast_fp16")];
            int32 var_3969 = const()[name = string("op_3969"), val = int32(1)];
            bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_85_cast_fp16 = concat(axis = var_3969, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_3971_cast_fp16))[name = string("doubled_85_cast_fp16")];
            tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881509376)))];
            fp16 var_3981_to_fp16 = const()[name = string("op_3981_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_3981_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_3992_split_sizes_0 = const()[name = string("op_3992_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3992_axis_0 = const()[name = string("op_3992_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_3992_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_3992_cast_fp16_1 = split(axis = var_3992_axis_0, split_sizes = var_3992_split_sizes_0, x = out_43_cast_fp16)[name = string("op_3992_cast_fp16")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_10_mlp_gate_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4009_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [2]> var_4015_strides_0 = const()[name = string("op_4015_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4015_pad_type_0 = const()[name = string("op_4015_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4015_pad_0 = const()[name = string("op_4015_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4015_dilations_0 = const()[name = string("op_4015_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4015_groups_0 = const()[name = string("op_4015_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4015_cast_fp16 = conv(dilations = var_4015_dilations_0, groups = var_4015_groups_0, pad = var_4015_pad_0, pad_type = var_4015_pad_type_0, strides = var_4015_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_3992_cast_fp16_0)[name = string("op_4015_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_109_cast_fp16 = mul(x = var_4009_cast_fp16, y = var_4015_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_107_pad_0 = const()[name = string("hidden_states_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_107_dilations_0 = const()[name = string("hidden_states_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_107_groups_0 = const()[name = string("hidden_states_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_107_cast_fp16 = conv(dilations = hidden_states_107_dilations_0, groups = hidden_states_107_groups_0, pad = hidden_states_107_pad_0, pad_type = hidden_states_107_pad_type_0, strides = hidden_states_107_strides_0, weight = layers_10_mlp_down_proj_weight_cast_fp16, x = x_109_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_109_cast_fp16 = add(x = hidden_states_105_cast_fp16, y = hidden_states_107_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4033_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_4033_cast_fp16")];
            int32 var_4031 = const()[name = string("op_4031"), val = int32(1)];
            bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_89_cast_fp16 = concat(axis = var_4031, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4033_cast_fp16))[name = string("doubled_89_cast_fp16")];
            tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881517632)))];
            fp16 var_4043_to_fp16 = const()[name = string("op_4043_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_4043_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4054_split_sizes_0 = const()[name = string("op_4054_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4054_axis_0 = const()[name = string("op_4054_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4054_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4054_cast_fp16_1 = split(axis = var_4054_axis_0, split_sizes = var_4054_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_67_pad_0 = const()[name = string("query_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_67_dilations_0 = const()[name = string("query_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_67_groups_0 = const()[name = string("query_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_67_cast_fp16 = conv(dilations = query_states_67_dilations_0, groups = query_states_67_groups_0, pad = query_states_67_pad_0, pad_type = query_states_67_pad_type_0, strides = query_states_67_strides_0, weight = layers_11_self_attn_q_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_111_pad_0 = const()[name = string("key_states_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_111_dilations_0 = const()[name = string("key_states_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_111_groups_0 = const()[name = string("key_states_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_111_cast_fp16 = conv(dilations = key_states_111_dilations_0, groups = key_states_111_groups_0, pad = key_states_111_pad_0, pad_type = key_states_111_pad_type_0, strides = key_states_111_strides_0, weight = layers_11_self_attn_k_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_67_pad_0 = const()[name = string("value_states_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_67_dilations_0 = const()[name = string("value_states_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_67_groups_0 = const()[name = string("value_states_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_67_cast_fp16 = conv(dilations = value_states_67_dilations_0, groups = value_states_67_groups_0, pad = value_states_67_pad_0, pad_type = value_states_67_pad_type_0, strides = value_states_67_strides_0, weight = layers_11_self_attn_v_proj_weight_cast_fp16, x = var_4054_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_111_cast_fp16 = reshape(shape = concat_132x, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4111_cast_fp16 = reshape(shape = concat_133x, x = key_states_111_cast_fp16)[name = string("op_4111_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4118_cast_fp16 = reshape(shape = concat_134x, x = value_states_67_cast_fp16)[name = string("op_4118_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4122_cast_fp16 = mul(x = x_111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4122_cast_fp16")];
            tensor<int32, [2]> var_4123_split_sizes_0 = const()[name = string("op_4123_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4123_axis_0 = const()[name = string("op_4123_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_4123_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_4123_cast_fp16_1 = split(axis = var_4123_axis_0, split_sizes = var_4123_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4123_cast_fp16")];
            fp16 const_114_promoted_to_fp16 = const()[name = string("const_114_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_4125_cast_fp16 = mul(x = var_4123_cast_fp16_1, y = const_114_promoted_to_fp16)[name = string("op_4125_cast_fp16")];
            int32 var_4127 = const()[name = string("op_4127"), val = int32(-2)];
            bool var_4128_interleave_0 = const()[name = string("op_4128_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_4128_cast_fp16 = concat(axis = var_4127, interleave = var_4128_interleave_0, values = (var_4125_cast_fp16, var_4123_cast_fp16_0))[name = string("op_4128_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4129_cast_fp16 = mul(x = var_4128_cast_fp16, y = var_459_cast_fp16)[name = string("op_4129_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_69_cast_fp16 = add(x = var_4122_cast_fp16, y = var_4129_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4135_cast_fp16 = mul(x = var_4111_cast_fp16, y = var_452_cast_fp16)[name = string("op_4135_cast_fp16")];
            tensor<int32, [2]> var_4136_split_sizes_0 = const()[name = string("op_4136_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4136_axis_0 = const()[name = string("op_4136_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_4136_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_4136_cast_fp16_1 = split(axis = var_4136_axis_0, split_sizes = var_4136_split_sizes_0, x = var_4111_cast_fp16)[name = string("op_4136_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_4138_cast_fp16 = mul(x = var_4136_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_4138_cast_fp16")];
            int32 var_4140 = const()[name = string("op_4140"), val = int32(-2)];
            bool var_4141_interleave_0 = const()[name = string("op_4141_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_4141_cast_fp16 = concat(axis = var_4140, interleave = var_4141_interleave_0, values = (var_4138_cast_fp16, var_4136_cast_fp16_0))[name = string("op_4141_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4142_cast_fp16 = mul(x = var_4141_cast_fp16, y = var_459_cast_fp16)[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_115_cast_fp16 = add(x = var_4135_cast_fp16, y = var_4142_cast_fp16)[name = string("key_states_115_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (expand_dims_132, expand_dims_133, position_id, expand_dims_135))[name = string("concat_137")];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> concat_138_values1_0 = const()[name = string("concat_138_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values3_0 = const()[name = string("concat_138_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (expand_dims_136, concat_138_values1_0, cache_position_end, concat_138_values3_0))[name = string("concat_138")];
            tensor<int32, [4]> key_states_117_perm_0 = const()[name = string("key_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_117_cast_fp16 = transpose(perm = key_states_117_perm_0, x = key_states_115_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = key_states_117_cast_fp16, x = coreml_update_state_20)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_22_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_22 = read_state(input = key_cache)[name = string("coreml_update_state_22")];
            tensor<int32, [4]> value_states_69_perm_0 = const()[name = string("value_states_69_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_69_cast_fp16 = transpose(perm = value_states_69_perm_0, x = var_4118_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_137, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_138, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = value_states_69_cast_fp16, x = coreml_update_state_21)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_23_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_23 = read_state(input = value_cache)[name = string("coreml_update_state_23")];
            tensor<int32, [4]> var_4212_begin_0 = const()[name = string("op_4212_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4212_end_0 = const()[name = string("op_4212_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4212_end_mask_0 = const()[name = string("op_4212_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4212_cast_fp16 = slice_by_index(begin = var_4212_begin_0, end = var_4212_end_0, end_mask = var_4212_end_mask_0, x = coreml_update_state_22)[name = string("op_4212_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4215_axis_0 = const()[name = string("op_4215_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4215_cast_fp16_1 = split(axis = var_4215_axis_0, split_sizes = tile_22, x = var_4212_cast_fp16)[name = string("op_4215_cast_fp16")];
            tensor<int32, [4]> var_4222_begin_0 = const()[name = string("op_4222_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4222_end_0 = const()[name = string("op_4222_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4222_end_mask_0 = const()[name = string("op_4222_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4222_cast_fp16 = slice_by_index(begin = var_4222_begin_0, end = var_4222_end_0, end_mask = var_4222_end_mask_0, x = coreml_update_state_23)[name = string("op_4222_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4225_axis_0 = const()[name = string("op_4225_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4225_cast_fp16_1 = split(axis = var_4225_axis_0, split_sizes = tile_23, x = var_4222_cast_fp16)[name = string("op_4225_cast_fp16")];
            tensor<int32, [2]> var_4228_split_sizes_0 = const()[name = string("op_4228_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4228_axis_0 = const()[name = string("op_4228_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_4228_0, tensor<fp16, [1, 8, 128, 1]> var_4228_1 = split(axis = var_4228_axis_0, split_sizes = var_4228_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4228")];
            bool attn_weights_177_transpose_x_0 = const()[name = string("attn_weights_177_transpose_x_0"), val = bool(false)];
            bool attn_weights_177_transpose_y_0 = const()[name = string("attn_weights_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_177_cast_fp16 = matmul(transpose_x = attn_weights_177_transpose_x_0, transpose_y = attn_weights_177_transpose_y_0, x = var_4215_cast_fp16_0, y = var_4228_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4231_to_fp16 = const()[name = string("op_4231_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_179_cast_fp16 = mul(x = attn_weights_177_cast_fp16, y = var_4231_to_fp16)[name = string("attn_weights_179_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_181_cast_fp16 = add(x = attn_weights_179_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_181_cast_fp16")];
            int32 var_4235 = const()[name = string("op_4235"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_183_cast_fp16 = softmax(axis = var_4235, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4241_transpose_x_1 = const()[name = string("op_4241_transpose_x_1"), val = bool(true)];
            bool var_4241_transpose_y_1 = const()[name = string("op_4241_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_4241_cast_fp16 = matmul(transpose_x = var_4241_transpose_x_1, transpose_y = var_4241_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4225_cast_fp16_0)[name = string("op_4241_cast_fp16")];
            bool attn_weights_185_transpose_x_0 = const()[name = string("attn_weights_185_transpose_x_0"), val = bool(false)];
            bool attn_weights_185_transpose_y_0 = const()[name = string("attn_weights_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_185_cast_fp16 = matmul(transpose_x = attn_weights_185_transpose_x_0, transpose_y = attn_weights_185_transpose_y_0, x = var_4215_cast_fp16_1, y = var_4228_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4243_to_fp16 = const()[name = string("op_4243_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_187_cast_fp16 = mul(x = attn_weights_185_cast_fp16, y = var_4243_to_fp16)[name = string("attn_weights_187_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_189_cast_fp16 = add(x = attn_weights_187_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_189_cast_fp16")];
            int32 var_4247 = const()[name = string("op_4247"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_191_cast_fp16 = softmax(axis = var_4247, x = attn_weights_189_cast_fp16)[name = string("attn_weights_191_cast_fp16")];
            bool attn_output_89_transpose_x_1 = const()[name = string("attn_output_89_transpose_x_1"), val = bool(true)];
            bool attn_output_89_transpose_y_1 = const()[name = string("attn_output_89_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_1, transpose_y = attn_output_89_transpose_y_1, x = attn_weights_191_cast_fp16, y = var_4225_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4255 = const()[name = string("op_4255"), val = int32(1)];
            bool attn_output_91_interleave_0 = const()[name = string("attn_output_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_91_cast_fp16 = concat(axis = var_4255, interleave = attn_output_91_interleave_0, values = (var_4241_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4259_perm_0 = const()[name = string("op_4259_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_4259_cast_fp16 = transpose(perm = var_4259_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_95_cast_fp16 = reshape(shape = concat_143x, x = var_4259_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_113_pad_0 = const()[name = string("hidden_states_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_113_dilations_0 = const()[name = string("hidden_states_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_113_groups_0 = const()[name = string("hidden_states_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_113_cast_fp16 = conv(dilations = hidden_states_113_dilations_0, groups = hidden_states_113_groups_0, pad = hidden_states_113_pad_0, pad_type = hidden_states_113_pad_type_0, strides = hidden_states_113_strides_0, weight = layers_11_self_attn_o_proj_weight_cast_fp16, x = attn_output_95_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_115_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = hidden_states_113_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4292_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_4292_cast_fp16")];
            int32 var_4290 = const()[name = string("op_4290"), val = int32(1)];
            bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_93_cast_fp16 = concat(axis = var_4290, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4292_cast_fp16))[name = string("doubled_93_cast_fp16")];
            tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881525888)))];
            fp16 var_4302_to_fp16 = const()[name = string("op_4302_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_4302_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4313_split_sizes_0 = const()[name = string("op_4313_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4313_axis_0 = const()[name = string("op_4313_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4313_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4313_cast_fp16_1 = split(axis = var_4313_axis_0, split_sizes = var_4313_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4313_cast_fp16")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_11_mlp_gate_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4330_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4330_cast_fp16")];
            tensor<int32, [2]> var_4336_strides_0 = const()[name = string("op_4336_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4336_pad_type_0 = const()[name = string("op_4336_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4336_pad_0 = const()[name = string("op_4336_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4336_dilations_0 = const()[name = string("op_4336_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4336_groups_0 = const()[name = string("op_4336_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4336_cast_fp16 = conv(dilations = var_4336_dilations_0, groups = var_4336_groups_0, pad = var_4336_pad_0, pad_type = var_4336_pad_type_0, strides = var_4336_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4313_cast_fp16_0)[name = string("op_4336_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_119_cast_fp16 = mul(x = var_4330_cast_fp16, y = var_4336_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_117_pad_0 = const()[name = string("hidden_states_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_117_dilations_0 = const()[name = string("hidden_states_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_117_groups_0 = const()[name = string("hidden_states_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_117_cast_fp16 = conv(dilations = hidden_states_117_dilations_0, groups = hidden_states_117_groups_0, pad = hidden_states_117_pad_0, pad_type = hidden_states_117_pad_type_0, strides = hidden_states_117_strides_0, weight = layers_11_mlp_down_proj_weight_cast_fp16, x = x_119_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_119_cast_fp16 = add(x = hidden_states_115_cast_fp16, y = hidden_states_117_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 const_122_promoted_to_fp16 = const()[name = string("const_122_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4354_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_122_promoted_to_fp16)[name = string("op_4354_cast_fp16")];
            int32 var_4352 = const()[name = string("op_4352"), val = int32(1)];
            bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_97_cast_fp16 = concat(axis = var_4352, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4354_cast_fp16))[name = string("doubled_97_cast_fp16")];
            tensor<int32, [1]> out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_49_gamma_0_to_fp16 = const()[name = string("out_49_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881534144)))];
            fp16 var_4364_to_fp16 = const()[name = string("op_4364_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_4364_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4375_split_sizes_0 = const()[name = string("op_4375_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4375_axis_0 = const()[name = string("op_4375_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4375_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4375_cast_fp16_1 = split(axis = var_4375_axis_0, split_sizes = var_4375_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4375_cast_fp16")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_73_pad_0 = const()[name = string("query_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_73_dilations_0 = const()[name = string("query_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_73_groups_0 = const()[name = string("query_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_73_cast_fp16 = conv(dilations = query_states_73_dilations_0, groups = query_states_73_groups_0, pad = query_states_73_pad_0, pad_type = query_states_73_pad_type_0, strides = query_states_73_strides_0, weight = layers_12_self_attn_q_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_121_pad_0 = const()[name = string("key_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_121_dilations_0 = const()[name = string("key_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_121_groups_0 = const()[name = string("key_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_121_cast_fp16 = conv(dilations = key_states_121_dilations_0, groups = key_states_121_groups_0, pad = key_states_121_pad_0, pad_type = key_states_121_pad_type_0, strides = key_states_121_strides_0, weight = layers_12_self_attn_k_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_73_pad_0 = const()[name = string("value_states_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_73_dilations_0 = const()[name = string("value_states_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_73_groups_0 = const()[name = string("value_states_73_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_73_cast_fp16 = conv(dilations = value_states_73_dilations_0, groups = value_states_73_groups_0, pad = value_states_73_pad_0, pad_type = value_states_73_pad_type_0, strides = value_states_73_strides_0, weight = layers_12_self_attn_v_proj_weight_cast_fp16, x = var_4375_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_121_cast_fp16 = reshape(shape = concat_144x, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4432_cast_fp16 = reshape(shape = concat_145x, x = key_states_121_cast_fp16)[name = string("op_4432_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4439_cast_fp16 = reshape(shape = concat_146x, x = value_states_73_cast_fp16)[name = string("op_4439_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4443_cast_fp16 = mul(x = x_121_cast_fp16, y = var_452_cast_fp16)[name = string("op_4443_cast_fp16")];
            tensor<int32, [2]> var_4444_split_sizes_0 = const()[name = string("op_4444_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4444_axis_0 = const()[name = string("op_4444_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_4444_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_4444_cast_fp16_1 = split(axis = var_4444_axis_0, split_sizes = var_4444_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4444_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_4446_cast_fp16 = mul(x = var_4444_cast_fp16_1, y = const_124_promoted_to_fp16)[name = string("op_4446_cast_fp16")];
            int32 var_4448 = const()[name = string("op_4448"), val = int32(-2)];
            bool var_4449_interleave_0 = const()[name = string("op_4449_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_4449_cast_fp16 = concat(axis = var_4448, interleave = var_4449_interleave_0, values = (var_4446_cast_fp16, var_4444_cast_fp16_0))[name = string("op_4449_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4450_cast_fp16 = mul(x = var_4449_cast_fp16, y = var_459_cast_fp16)[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_75_cast_fp16 = add(x = var_4443_cast_fp16, y = var_4450_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4456_cast_fp16 = mul(x = var_4432_cast_fp16, y = var_452_cast_fp16)[name = string("op_4456_cast_fp16")];
            tensor<int32, [2]> var_4457_split_sizes_0 = const()[name = string("op_4457_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4457_axis_0 = const()[name = string("op_4457_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_4457_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_4457_cast_fp16_1 = split(axis = var_4457_axis_0, split_sizes = var_4457_split_sizes_0, x = var_4432_cast_fp16)[name = string("op_4457_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_4459_cast_fp16 = mul(x = var_4457_cast_fp16_1, y = const_125_promoted_to_fp16)[name = string("op_4459_cast_fp16")];
            int32 var_4461 = const()[name = string("op_4461"), val = int32(-2)];
            bool var_4462_interleave_0 = const()[name = string("op_4462_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_4462_cast_fp16 = concat(axis = var_4461, interleave = var_4462_interleave_0, values = (var_4459_cast_fp16, var_4457_cast_fp16_0))[name = string("op_4462_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4463_cast_fp16 = mul(x = var_4462_cast_fp16, y = var_459_cast_fp16)[name = string("op_4463_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_125_cast_fp16 = add(x = var_4456_cast_fp16, y = var_4463_cast_fp16)[name = string("key_states_125_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (expand_dims_144, expand_dims_145, position_id, expand_dims_147))[name = string("concat_149")];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_150_values3_0 = const()[name = string("concat_150_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (expand_dims_148, concat_150_values1_0, cache_position_end, concat_150_values3_0))[name = string("concat_150")];
            tensor<int32, [4]> key_states_127_perm_0 = const()[name = string("key_states_127_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_127_cast_fp16 = transpose(perm = key_states_127_perm_0, x = key_states_125_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = key_states_127_cast_fp16, x = coreml_update_state_22)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_24_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_24 = read_state(input = key_cache)[name = string("coreml_update_state_24")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = string("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_4439_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_149, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_150, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = value_states_75_cast_fp16, x = coreml_update_state_23)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_25_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_25 = read_state(input = value_cache)[name = string("coreml_update_state_25")];
            tensor<int32, [4]> var_4533_begin_0 = const()[name = string("op_4533_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4533_end_0 = const()[name = string("op_4533_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4533_end_mask_0 = const()[name = string("op_4533_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4533_cast_fp16 = slice_by_index(begin = var_4533_begin_0, end = var_4533_end_0, end_mask = var_4533_end_mask_0, x = coreml_update_state_24)[name = string("op_4533_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4536_axis_0 = const()[name = string("op_4536_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4536_cast_fp16_1 = split(axis = var_4536_axis_0, split_sizes = tile_24, x = var_4533_cast_fp16)[name = string("op_4536_cast_fp16")];
            tensor<int32, [4]> var_4543_begin_0 = const()[name = string("op_4543_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4543_end_0 = const()[name = string("op_4543_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4543_end_mask_0 = const()[name = string("op_4543_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4543_cast_fp16 = slice_by_index(begin = var_4543_begin_0, end = var_4543_end_0, end_mask = var_4543_end_mask_0, x = coreml_update_state_25)[name = string("op_4543_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4546_axis_0 = const()[name = string("op_4546_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4546_cast_fp16_1 = split(axis = var_4546_axis_0, split_sizes = tile_25, x = var_4543_cast_fp16)[name = string("op_4546_cast_fp16")];
            tensor<int32, [2]> var_4549_split_sizes_0 = const()[name = string("op_4549_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4549_axis_0 = const()[name = string("op_4549_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_4549_0, tensor<fp16, [1, 8, 128, 1]> var_4549_1 = split(axis = var_4549_axis_0, split_sizes = var_4549_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4549")];
            bool attn_weights_193_transpose_x_0 = const()[name = string("attn_weights_193_transpose_x_0"), val = bool(false)];
            bool attn_weights_193_transpose_y_0 = const()[name = string("attn_weights_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_193_cast_fp16 = matmul(transpose_x = attn_weights_193_transpose_x_0, transpose_y = attn_weights_193_transpose_y_0, x = var_4536_cast_fp16_0, y = var_4549_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4552_to_fp16 = const()[name = string("op_4552_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_195_cast_fp16 = mul(x = attn_weights_193_cast_fp16, y = var_4552_to_fp16)[name = string("attn_weights_195_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_197_cast_fp16 = add(x = attn_weights_195_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_197_cast_fp16")];
            int32 var_4556 = const()[name = string("op_4556"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_199_cast_fp16 = softmax(axis = var_4556, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4562_transpose_x_1 = const()[name = string("op_4562_transpose_x_1"), val = bool(true)];
            bool var_4562_transpose_y_1 = const()[name = string("op_4562_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_4562_cast_fp16 = matmul(transpose_x = var_4562_transpose_x_1, transpose_y = var_4562_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4546_cast_fp16_0)[name = string("op_4562_cast_fp16")];
            bool attn_weights_201_transpose_x_0 = const()[name = string("attn_weights_201_transpose_x_0"), val = bool(false)];
            bool attn_weights_201_transpose_y_0 = const()[name = string("attn_weights_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_201_cast_fp16 = matmul(transpose_x = attn_weights_201_transpose_x_0, transpose_y = attn_weights_201_transpose_y_0, x = var_4536_cast_fp16_1, y = var_4549_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_203_cast_fp16 = mul(x = attn_weights_201_cast_fp16, y = var_4564_to_fp16)[name = string("attn_weights_203_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_205_cast_fp16 = add(x = attn_weights_203_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_205_cast_fp16")];
            int32 var_4568 = const()[name = string("op_4568"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_207_cast_fp16 = softmax(axis = var_4568, x = attn_weights_205_cast_fp16)[name = string("attn_weights_207_cast_fp16")];
            bool attn_output_97_transpose_x_1 = const()[name = string("attn_output_97_transpose_x_1"), val = bool(true)];
            bool attn_output_97_transpose_y_1 = const()[name = string("attn_output_97_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_1, transpose_y = attn_output_97_transpose_y_1, x = attn_weights_207_cast_fp16, y = var_4546_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4576 = const()[name = string("op_4576"), val = int32(1)];
            bool attn_output_99_interleave_0 = const()[name = string("attn_output_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_99_cast_fp16 = concat(axis = var_4576, interleave = attn_output_99_interleave_0, values = (var_4562_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4580_perm_0 = const()[name = string("op_4580_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_155x = const()[name = string("concat_155x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_4580_cast_fp16 = transpose(perm = var_4580_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_103_cast_fp16 = reshape(shape = concat_155x, x = var_4580_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_123_pad_0 = const()[name = string("hidden_states_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_123_dilations_0 = const()[name = string("hidden_states_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_123_groups_0 = const()[name = string("hidden_states_123_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_123_cast_fp16 = conv(dilations = hidden_states_123_dilations_0, groups = hidden_states_123_groups_0, pad = hidden_states_123_pad_0, pad_type = hidden_states_123_pad_type_0, strides = hidden_states_123_strides_0, weight = layers_12_self_attn_o_proj_weight_cast_fp16, x = attn_output_103_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_125_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_125_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4613_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4613_cast_fp16")];
            int32 var_4611 = const()[name = string("op_4611"), val = int32(1)];
            bool doubled_101_interleave_0 = const()[name = string("doubled_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_101_cast_fp16 = concat(axis = var_4611, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_4613_cast_fp16))[name = string("doubled_101_cast_fp16")];
            tensor<int32, [1]> out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_51_gamma_0_to_fp16 = const()[name = string("out_51_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881542400)))];
            fp16 var_4623_to_fp16 = const()[name = string("op_4623_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_4623_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_4634_split_sizes_0 = const()[name = string("op_4634_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4634_axis_0 = const()[name = string("op_4634_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4634_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4634_cast_fp16_1 = split(axis = var_4634_axis_0, split_sizes = var_4634_split_sizes_0, x = out_51_cast_fp16)[name = string("op_4634_cast_fp16")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_12_mlp_gate_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4651_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_4651_cast_fp16")];
            tensor<int32, [2]> var_4657_strides_0 = const()[name = string("op_4657_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4657_pad_type_0 = const()[name = string("op_4657_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4657_pad_0 = const()[name = string("op_4657_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4657_dilations_0 = const()[name = string("op_4657_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4657_groups_0 = const()[name = string("op_4657_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4657_cast_fp16 = conv(dilations = var_4657_dilations_0, groups = var_4657_groups_0, pad = var_4657_pad_0, pad_type = var_4657_pad_type_0, strides = var_4657_strides_0, weight = layers_12_mlp_up_proj_weight_cast_fp16, x = var_4634_cast_fp16_0)[name = string("op_4657_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_129_cast_fp16 = mul(x = var_4651_cast_fp16, y = var_4657_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_127_pad_0 = const()[name = string("hidden_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_127_dilations_0 = const()[name = string("hidden_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_127_groups_0 = const()[name = string("hidden_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_127_cast_fp16 = conv(dilations = hidden_states_127_dilations_0, groups = hidden_states_127_groups_0, pad = hidden_states_127_pad_0, pad_type = hidden_states_127_pad_type_0, strides = hidden_states_127_strides_0, weight = layers_12_mlp_down_proj_weight_cast_fp16, x = x_129_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_129_cast_fp16 = add(x = hidden_states_125_cast_fp16, y = hidden_states_127_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4675_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_4675_cast_fp16")];
            int32 var_4673 = const()[name = string("op_4673"), val = int32(1)];
            bool doubled_105_interleave_0 = const()[name = string("doubled_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_105_cast_fp16 = concat(axis = var_4673, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_4675_cast_fp16))[name = string("doubled_105_cast_fp16")];
            tensor<int32, [1]> out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_53_gamma_0_to_fp16 = const()[name = string("out_53_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881550656)))];
            fp16 var_4685_to_fp16 = const()[name = string("op_4685_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_4685_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_4696_split_sizes_0 = const()[name = string("op_4696_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4696_axis_0 = const()[name = string("op_4696_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4696_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4696_cast_fp16_1 = split(axis = var_4696_axis_0, split_sizes = var_4696_split_sizes_0, x = out_53_cast_fp16)[name = string("op_4696_cast_fp16")];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> query_states_79_pad_0 = const()[name = string("query_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_79_dilations_0 = const()[name = string("query_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_79_groups_0 = const()[name = string("query_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> query_states_79_cast_fp16 = conv(dilations = query_states_79_dilations_0, groups = query_states_79_groups_0, pad = query_states_79_pad_0, pad_type = query_states_79_pad_type_0, strides = query_states_79_strides_0, weight = layers_13_self_attn_q_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> key_states_131_pad_0 = const()[name = string("key_states_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_131_dilations_0 = const()[name = string("key_states_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_131_groups_0 = const()[name = string("key_states_131_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> key_states_131_cast_fp16 = conv(dilations = key_states_131_dilations_0, groups = key_states_131_groups_0, pad = key_states_131_pad_0, pad_type = key_states_131_pad_type_0, strides = key_states_131_strides_0, weight = layers_13_self_attn_k_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> value_states_79_pad_0 = const()[name = string("value_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_79_dilations_0 = const()[name = string("value_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_79_groups_0 = const()[name = string("value_states_79_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 1]> value_states_79_cast_fp16 = conv(dilations = value_states_79_dilations_0, groups = value_states_79_groups_0, pad = value_states_79_pad_0, pad_type = value_states_79_pad_type_0, strides = value_states_79_strides_0, weight = layers_13_self_attn_v_proj_weight_cast_fp16, x = var_4696_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, 16, 128, -1])];
            tensor<fp16, [1, 16, 128, 1]> x_131_cast_fp16 = reshape(shape = concat_156x, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4753_cast_fp16 = reshape(shape = concat_157x, x = key_states_131_cast_fp16)[name = string("op_4753_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, 2, 128, -1])];
            tensor<fp16, [1, 2, 128, 1]> var_4760_cast_fp16 = reshape(shape = concat_158x, x = value_states_79_cast_fp16)[name = string("op_4760_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4764_cast_fp16 = mul(x = x_131_cast_fp16, y = var_452_cast_fp16)[name = string("op_4764_cast_fp16")];
            tensor<int32, [2]> var_4765_split_sizes_0 = const()[name = string("op_4765_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4765_axis_0 = const()[name = string("op_4765_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 1]> var_4765_cast_fp16_0, tensor<fp16, [1, 16, 64, 1]> var_4765_cast_fp16_1 = split(axis = var_4765_axis_0, split_sizes = var_4765_split_sizes_0, x = x_131_cast_fp16)[name = string("op_4765_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 1]> var_4767_cast_fp16 = mul(x = var_4765_cast_fp16_1, y = const_134_promoted_to_fp16)[name = string("op_4767_cast_fp16")];
            int32 var_4769 = const()[name = string("op_4769"), val = int32(-2)];
            bool var_4770_interleave_0 = const()[name = string("op_4770_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 1]> var_4770_cast_fp16 = concat(axis = var_4769, interleave = var_4770_interleave_0, values = (var_4767_cast_fp16, var_4765_cast_fp16_0))[name = string("op_4770_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> var_4771_cast_fp16 = mul(x = var_4770_cast_fp16, y = var_459_cast_fp16)[name = string("op_4771_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1]> query_states_81_cast_fp16 = add(x = var_4764_cast_fp16, y = var_4771_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4777_cast_fp16 = mul(x = var_4753_cast_fp16, y = var_452_cast_fp16)[name = string("op_4777_cast_fp16")];
            tensor<int32, [2]> var_4778_split_sizes_0 = const()[name = string("op_4778_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4778_axis_0 = const()[name = string("op_4778_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 1]> var_4778_cast_fp16_0, tensor<fp16, [1, 2, 64, 1]> var_4778_cast_fp16_1 = split(axis = var_4778_axis_0, split_sizes = var_4778_split_sizes_0, x = var_4753_cast_fp16)[name = string("op_4778_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 1]> var_4780_cast_fp16 = mul(x = var_4778_cast_fp16_1, y = const_135_promoted_to_fp16)[name = string("op_4780_cast_fp16")];
            int32 var_4782 = const()[name = string("op_4782"), val = int32(-2)];
            bool var_4783_interleave_0 = const()[name = string("op_4783_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 1]> var_4783_cast_fp16 = concat(axis = var_4782, interleave = var_4783_interleave_0, values = (var_4780_cast_fp16, var_4778_cast_fp16_0))[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4784_cast_fp16 = mul(x = var_4783_cast_fp16, y = var_459_cast_fp16)[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> key_states_135_cast_fp16 = add(x = var_4777_cast_fp16, y = var_4784_cast_fp16)[name = string("key_states_135_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (expand_dims_156, expand_dims_157, position_id, expand_dims_159))[name = string("concat_161")];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> concat_162_values1_0 = const()[name = string("concat_162_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_162_values3_0 = const()[name = string("concat_162_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_162_axis_0 = const()[name = string("concat_162_axis_0"), val = int32(0)];
            bool concat_162_interleave_0 = const()[name = string("concat_162_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_162 = concat(axis = concat_162_axis_0, interleave = concat_162_interleave_0, values = (expand_dims_160, concat_162_values1_0, cache_position_end, concat_162_values3_0))[name = string("concat_162")];
            tensor<int32, [4]> key_states_137_perm_0 = const()[name = string("key_states_137_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> key_states_137_cast_fp16 = transpose(perm = key_states_137_perm_0, x = key_states_135_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [14, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = key_states_137_cast_fp16, x = coreml_update_state_24)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_26_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_26 = read_state(input = key_cache)[name = string("coreml_update_state_26")];
            tensor<int32, [4]> value_states_81_perm_0 = const()[name = string("value_states_81_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 1, 128]> value_states_81_cast_fp16 = transpose(perm = value_states_81_perm_0, x = var_4760_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [14, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_161, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_162, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = value_states_81_cast_fp16, x = coreml_update_state_25)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_27_write_state")];
            tensor<fp16, [14, 2, 2048, 128]> coreml_update_state_27 = read_state(input = value_cache)[name = string("coreml_update_state_27")];
            tensor<int32, [4]> var_4854_begin_0 = const()[name = string("op_4854_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4854_end_0 = const()[name = string("op_4854_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4854_end_mask_0 = const()[name = string("op_4854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4854_cast_fp16 = slice_by_index(begin = var_4854_begin_0, end = var_4854_end_0, end_mask = var_4854_end_mask_0, x = coreml_update_state_26)[name = string("op_4854_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4857_axis_0 = const()[name = string("op_4857_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4857_cast_fp16_1 = split(axis = var_4857_axis_0, split_sizes = tile_26, x = var_4854_cast_fp16)[name = string("op_4857_cast_fp16")];
            tensor<int32, [4]> var_4864_begin_0 = const()[name = string("op_4864_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_4864_end_0 = const()[name = string("op_4864_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_4864_end_mask_0 = const()[name = string("op_4864_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4864_cast_fp16 = slice_by_index(begin = var_4864_begin_0, end = var_4864_end_0, end_mask = var_4864_end_mask_0, x = coreml_update_state_27)[name = string("op_4864_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4867_axis_0 = const()[name = string("op_4867_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4867_cast_fp16_1 = split(axis = var_4867_axis_0, split_sizes = tile_27, x = var_4864_cast_fp16)[name = string("op_4867_cast_fp16")];
            tensor<int32, [2]> var_4870_split_sizes_0 = const()[name = string("op_4870_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4870_axis_0 = const()[name = string("op_4870_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 1]> var_4870_0, tensor<fp16, [1, 8, 128, 1]> var_4870_1 = split(axis = var_4870_axis_0, split_sizes = var_4870_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_4870")];
            bool attn_weights_209_transpose_x_0 = const()[name = string("attn_weights_209_transpose_x_0"), val = bool(false)];
            bool attn_weights_209_transpose_y_0 = const()[name = string("attn_weights_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_209_cast_fp16 = matmul(transpose_x = attn_weights_209_transpose_x_0, transpose_y = attn_weights_209_transpose_y_0, x = var_4857_cast_fp16_0, y = var_4870_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_4873_to_fp16 = const()[name = string("op_4873_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_211_cast_fp16 = mul(x = attn_weights_209_cast_fp16, y = var_4873_to_fp16)[name = string("attn_weights_211_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_213_cast_fp16 = add(x = attn_weights_211_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_213_cast_fp16")];
            int32 var_4877 = const()[name = string("op_4877"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_215_cast_fp16 = softmax(axis = var_4877, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_4883_transpose_x_1 = const()[name = string("op_4883_transpose_x_1"), val = bool(true)];
            bool var_4883_transpose_y_1 = const()[name = string("op_4883_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> var_4883_cast_fp16 = matmul(transpose_x = var_4883_transpose_x_1, transpose_y = var_4883_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_4867_cast_fp16_0)[name = string("op_4883_cast_fp16")];
            bool attn_weights_217_transpose_x_0 = const()[name = string("attn_weights_217_transpose_x_0"), val = bool(false)];
            bool attn_weights_217_transpose_y_0 = const()[name = string("attn_weights_217_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_217_cast_fp16 = matmul(transpose_x = attn_weights_217_transpose_x_0, transpose_y = attn_weights_217_transpose_y_0, x = var_4857_cast_fp16_1, y = var_4870_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_4885_to_fp16 = const()[name = string("op_4885_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_219_cast_fp16 = mul(x = attn_weights_217_cast_fp16, y = var_4885_to_fp16)[name = string("attn_weights_219_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_221_cast_fp16 = add(x = attn_weights_219_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_221_cast_fp16")];
            int32 var_4889 = const()[name = string("op_4889"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 1]> attn_weights_cast_fp16 = softmax(axis = var_4889, x = attn_weights_221_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_105_transpose_x_1 = const()[name = string("attn_output_105_transpose_x_1"), val = bool(true)];
            bool attn_output_105_transpose_y_1 = const()[name = string("attn_output_105_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 1, 128]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_1, transpose_y = attn_output_105_transpose_y_1, x = attn_weights_cast_fp16, y = var_4867_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_4897 = const()[name = string("op_4897"), val = int32(1)];
            bool attn_output_107_interleave_0 = const()[name = string("attn_output_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1, 128]> attn_output_107_cast_fp16 = concat(axis = var_4897, interleave = attn_output_107_interleave_0, values = (var_4883_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_4901_perm_0 = const()[name = string("op_4901_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [4]>([1, 2048, 1, -1])];
            tensor<fp16, [1, 16, 128, 1]> var_4901_cast_fp16 = transpose(perm = var_4901_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 2048, 1, 1]> attn_output_cast_fp16 = reshape(shape = concat_167x, x = var_4901_cast_fp16)[name = string("attn_output_cast_fp16")];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_133_pad_0 = const()[name = string("hidden_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_133_dilations_0 = const()[name = string("hidden_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_133_groups_0 = const()[name = string("hidden_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_133_cast_fp16 = conv(dilations = hidden_states_133_dilations_0, groups = hidden_states_133_groups_0, pad = hidden_states_133_pad_0, pad_type = hidden_states_133_pad_type_0, strides = hidden_states_133_strides_0, weight = layers_13_self_attn_o_proj_weight_cast_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_133_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_135_cast_fp16 = add(x = hidden_states_129_cast_fp16, y = hidden_states_133_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 1]> var_4934_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_4934_cast_fp16")];
            int32 var_4932 = const()[name = string("op_4932"), val = int32(1)];
            bool doubled_109_interleave_0 = const()[name = string("doubled_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 1]> doubled_109_cast_fp16 = concat(axis = var_4932, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_4934_cast_fp16))[name = string("doubled_109_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(881558912)))];
            fp16 var_4944_to_fp16 = const()[name = string("op_4944_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_4944_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_4955_split_sizes_0 = const()[name = string("op_4955_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4955_axis_0 = const()[name = string("op_4955_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> var_4955_cast_fp16_0, tensor<fp16, [1, 2048, 1, 1]> var_4955_cast_fp16_1 = split(axis = var_4955_axis_0, split_sizes = var_4955_split_sizes_0, x = out_cast_fp16)[name = string("op_4955_cast_fp16")];
            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_13_mlp_gate_proj_weight_cast_fp16, x = var_4955_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> var_4972_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_4972_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(881567168)))];
            tensor<int32, [2]> var_4978_strides_0 = const()[name = string("op_4978_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string var_4978_pad_type_0 = const()[name = string("op_4978_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> var_4978_pad_0 = const()[name = string("op_4978_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4978_dilations_0 = const()[name = string("op_4978_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4978_groups_0 = const()[name = string("op_4978_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 1]> var_4978_cast_fp16 = conv(dilations = var_4978_dilations_0, groups = var_4978_groups_0, pad = var_4978_pad_0, pad_type = var_4978_pad_type_0, strides = var_4978_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_4955_cast_fp16_0)[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 1]> x_cast_fp16 = mul(x = var_4972_cast_fp16, y = var_4978_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [2]> hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor<int32, [2]>([1, 1])];
            string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")];
            tensor<int32, [4]> hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states_cast_fp16 = conv(dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_13_mlp_down_proj_weight_cast_fp16, x = x_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 1]> hidden_states = add(x = hidden_states_135_cast_fp16, y = hidden_states_cast_fp16)[name = string("op_4987_cast_fp16")];
        } -> (hidden_states);
}