program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.11.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios18>(tensor<fp32, [1, 2048, 1, 8]> inputs_embeds, state<tensor<fp16, [28, 2, 2048, 128]>> key_cache, tensor<int32, [1]> position_id, tensor<int32, [8]> position_index_seed, state<tensor<fp16, [28, 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")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_14_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(432534656))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436733184))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436729024))))[name = string("layers_14_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_14_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(436735296))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437260224))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437259648))))[name = string("layers_14_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_14_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(437260544))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441459072))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441454912))))[name = string("layers_14_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_14_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(441461184))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454056512))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454044160))))[name = string("layers_14_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_14_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(454062720))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466658048))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466645696))))[name = string("layers_14_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_14_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(466664256))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(479251392))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(479247232))))[name = string("layers_14_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_15_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(479253504))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483452032))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483447872))))[name = string("layers_15_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_15_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(483454144))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483979072))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483978496))))[name = string("layers_15_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_15_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(483979392))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484504320))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484503744))))[name = string("layers_15_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_15_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(484504640))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488703168))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488699008))))[name = string("layers_15_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_15_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(488705280))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501300608))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501288256))))[name = string("layers_15_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_15_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(501306816))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513902144))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513889792))))[name = string("layers_15_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_15_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(513908352))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526495488))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526491328))))[name = string("layers_15_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_16_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(526497600))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530696128))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530691968))))[name = string("layers_16_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_16_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(530698240))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531223168))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531222592))))[name = string("layers_16_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_16_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(531223488))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531748416))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531747840))))[name = string("layers_16_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_16_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(531748736))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535947264))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535943104))))[name = string("layers_16_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_16_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(535949376))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(548536512))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(548532352))))[name = string("layers_16_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_17_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(548538624))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552737152))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552732992))))[name = string("layers_17_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_17_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(552739264))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553264192))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553263616))))[name = string("layers_17_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_17_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(553264512))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553789440))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553788864))))[name = string("layers_17_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_17_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(553789760))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557988288))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557984128))))[name = string("layers_17_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_17_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(557990400))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570585728))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570573376))))[name = string("layers_17_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_17_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(570591936))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583187264))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583174912))))[name = string("layers_17_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_17_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(583193472))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595780608))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595776448))))[name = string("layers_17_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_18_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(595782720))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599981248))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599977088))))[name = string("layers_18_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_18_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(599983360))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600508288))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600507712))))[name = string("layers_18_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_18_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(600508608))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601033536))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601032960))))[name = string("layers_18_self_attn_v_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_18_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(601033856))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605232384))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605228224))))[name = string("layers_18_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_18_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(605234496))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(617829824))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(617817472))))[name = string("layers_18_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_18_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(617836032))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630431360))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630419008))))[name = string("layers_18_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_18_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(630437568))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643024704))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643020544))))[name = string("layers_18_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_19_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(643026816))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647225344))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647221184))))[name = string("layers_19_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_19_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(647227456))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647752384))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647751808))))[name = string("layers_19_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_19_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(647752704))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660348032))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660335680))))[name = string("layers_19_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_19_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(660354240))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(672949568))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(672937216))))[name = string("layers_19_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_19_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(672955776))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685542912))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685538752))))[name = string("layers_19_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_20_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(685545024))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689743552))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689739392))))[name = string("layers_20_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_20_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(689745664))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690270592))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690270016))))[name = string("layers_20_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_20_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(690270912))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694469440))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694465280))))[name = string("layers_20_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_20_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(694471552))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(707066880))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(707054528))))[name = string("layers_20_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_20_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(707073088))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719660224))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719656064))))[name = string("layers_20_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_21_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(719662336))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723860864))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723856704))))[name = string("layers_21_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_21_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(723862976))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724387904))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724387328))))[name = string("layers_21_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_21_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(724388224))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(728586752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(728582592))))[name = string("layers_21_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_21_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(728588864))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(741184192))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(741171840))))[name = string("layers_21_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_21_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(741190400))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(753785728))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(753773376))))[name = string("layers_21_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_21_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(753791936))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(766379072))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(766374912))))[name = string("layers_21_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_22_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(766381184))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(770579712))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(770575552))))[name = string("layers_22_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_22_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(770581824))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(771106752))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(771106176))))[name = string("layers_22_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_22_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(771107072))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(783702400))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(783690048))))[name = string("layers_22_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_22_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(783708608))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(796303936))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(796291584))))[name = string("layers_22_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_22_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(796310144))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(808897280))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(808893120))))[name = string("layers_22_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_23_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(808899392))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(813097920))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(813093760))))[name = string("layers_23_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_23_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(813100032))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(813624960))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(813624384))))[name = string("layers_23_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_23_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(813625280))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817823808))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817819648))))[name = string("layers_23_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_23_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(817825920))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(830421248))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(830408896))))[name = string("layers_23_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_23_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(830427456))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(843022784))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(843010432))))[name = string("layers_23_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_23_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(843028992))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(855616128))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(855611968))))[name = string("layers_23_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_24_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(855618240))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(859816768))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(859812608))))[name = string("layers_24_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_24_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(859818880))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(860343808))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(860343232))))[name = string("layers_24_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_24_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(860344128))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(864542656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(864538496))))[name = string("layers_24_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_24_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(864544768))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(877140096))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(877127744))))[name = string("layers_24_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_24_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(877146304))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(889741632))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(889729280))))[name = string("layers_24_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 6144, 1, 1]> layers_24_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(889747840))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(902334976))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(902330816))))[name = string("layers_24_mlp_down_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_25_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(902337088))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(906535616))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(906531456))))[name = string("layers_25_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_25_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(906537728))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(907062656))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(907062080))))[name = string("layers_25_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_25_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(907062976))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(911261504))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(911257344))))[name = string("layers_25_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_25_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(911263616))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(923858944))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(923846592))))[name = string("layers_25_mlp_gate_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_25_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(923865152))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(936460480))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(936448128))))[name = string("layers_25_mlp_up_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_26_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(936466688))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(940665216))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(940661056))))[name = string("layers_26_self_attn_q_proj_weight_cast_fp16")];
            tensor<fp16, [256, 2048, 1, 1]> layers_26_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(940667328))), offset = tensor<int8, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(941192256))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(941191680))))[name = string("layers_26_self_attn_k_proj_weight_cast_fp16")];
            tensor<fp16, [2048, 2048, 1, 1]> layers_26_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(941192576))), offset = tensor<int8, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(945391104))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(945386944))))[name = string("layers_26_self_attn_o_proj_weight_cast_fp16")];
            tensor<fp16, [6144, 2048, 1, 1]> layers_26_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(945393216))), offset = tensor<int8, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(957988544))), scale = tensor<fp16, [6144, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(957976192))))[name = string("layers_26_mlp_gate_proj_weight_cast_fp16")];
            tensor<int32, [1]> var_766 = const()[name = string("op_766"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> cache_position_end = add(x = position_id, y = var_766)[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_831 = const()[name = string("op_831"), val = int32(0)];
            bool var_833_exclusive_0 = const()[name = string("op_833_exclusive_0"), val = bool(false)];
            bool var_833_reverse_0 = const()[name = string("op_833_reverse_0"), val = bool(false)];
            tensor<fp16, [8]> var_833_cast_fp16 = cumsum(axis = var_831, exclusive = var_833_exclusive_0, reverse = var_833_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_833_cast_fp16")];
            fp16 var_835_promoted_to_fp16 = const()[name = string("op_835_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [8]> position_offsets_cast_fp16 = sub(x = var_833_cast_fp16, y = var_835_promoted_to_fp16)[name = string("position_offsets_cast_fp16")];
            tensor<int32, [1]> var_838_axes_0 = const()[name = string("op_838_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8]> var_838_cast_fp16 = expand_dims(axes = var_838_axes_0, x = position_offsets_cast_fp16)[name = string("op_838_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_231")];
            tensor<fp16, [1, 8]> position_ids_1_cast_fp16 = add(x = var_838_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 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)];
            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(957994752)))];
            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_230")];
            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")];
            int32 cos_1_cast_fp16_axis_0 = const()[name = string("cos_1_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 8, 128]> cos_1_cast_fp16 = gather(axis = cos_1_cast_fp16_axis_0, batch_dims = cos_1_batch_dims_0, indices = select_0, validate_indices = cos_1_validate_indices_0, x = rope_emb_cos_cached_to_fp16)[name = string("cos_1_cast_fp16")];
            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)];
            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(966383424)))];
            int32 sin_1_cast_fp16_axis_0 = const()[name = string("sin_1_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 8, 128]> sin_1_cast_fp16 = gather(axis = sin_1_cast_fp16_axis_0, batch_dims = sin_1_batch_dims_0, indices = select_0, validate_indices = sin_1_validate_indices_0, x = rope_emb_sin_cached_to_fp16)[name = string("sin_1_cast_fp16")];
            tensor<int32, [3]> var_857_perm_0 = const()[name = string("op_857_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_859_axes_0 = const()[name = string("op_859_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 8]> var_857_cast_fp16 = transpose(perm = var_857_perm_0, x = cos_1_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 1, 128, 8]> var_859_cast_fp16 = expand_dims(axes = var_859_axes_0, x = var_857_cast_fp16)[name = string("op_859_cast_fp16")];
            tensor<int32, [3]> var_864_perm_0 = const()[name = string("op_864_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
            tensor<int32, [1]> var_866_axes_0 = const()[name = string("op_866_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 8]> var_864_cast_fp16 = transpose(perm = var_864_perm_0, x = sin_1_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 1, 128, 8]> var_866_cast_fp16 = expand_dims(axes = var_866_axes_0, x = var_864_cast_fp16)[name = string("op_866_cast_fp16")];
            tensor<int32, [1, 1, 2048]> var_878 = const()[name = string("op_878"), val = tensor<int32, [1, 1, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(974772096)))];
            tensor<int32, [1]> var_885_axes_0 = const()[name = string("op_885_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 8, 1]> var_885 = expand_dims(axes = var_885_axes_0, x = position_ids_1_cast_fp16_to_int32)[name = string("op_885")];
            tensor<bool, [1, 8, 2048]> var_886 = greater(x = var_878, y = var_885)[name = string("op_886")];
            tensor<int32, [1]> var_893_axes_0 = const()[name = string("op_893_axes_0"), val = tensor<int32, [1]>([1])];
            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 8, 2048]> var_886_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_886)[name = string("cast_229")];
            tensor<fp16, [1, 1, 8, 2048]> var_893_cast_fp16 = expand_dims(axes = var_893_axes_0, x = var_886_to_fp16)[name = string("op_893_cast_fp16")];
            tensor<int32, [4]> mask_perm_0 = const()[name = string("mask_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_897_promoted_to_fp16 = const()[name = string("op_897_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 2048, 8]> mask_cast_fp16 = transpose(perm = mask_perm_0, x = var_893_cast_fp16)[name = string("transpose_84")];
            tensor<bool, [1, 1, 2048, 8]> var_898_cast_fp16 = equal(x = mask_cast_fp16, y = var_897_promoted_to_fp16)[name = string("op_898_cast_fp16")];
            tensor<fp16, [1, 1, 2048, 8]> var_899_after_broadcast_to_fp16 = const()[name = string("op_899_after_broadcast_to_fp16"), val = tensor<fp16, [1, 1, 2048, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(974780352)))];
            tensor<fp16, [1, 1, 2048, 8]> attn_mask_1_cast_fp16 = select(a = mask_cast_fp16, b = var_899_after_broadcast_to_fp16, cond = var_898_cast_fp16)[name = string("attn_mask_1_cast_fp16")];
            int32 var_907 = const()[name = string("op_907"), val = int32(1)];
            string inputs_embeds_to_fp16_dtype_0 = const()[name = string("inputs_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_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_228")];
            tensor<fp16, [1, 2048, 1, 8]> var_909_cast_fp16 = mul(x = inputs_embeds_to_fp16, y = const_3_promoted_to_fp16)[name = string("op_909_cast_fp16")];
            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_907, interleave = doubled_1_interleave_0, values = (inputs_embeds_to_fp16, var_909_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(974813184)))];
            fp16 var_919_to_fp16 = const()[name = string("op_919_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_919_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [2]> var_930_split_sizes_0 = const()[name = string("op_930_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_930_axis_0 = const()[name = string("op_930_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_930_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_930_cast_fp16_1 = split(axis = var_930_axis_0, split_sizes = var_930_split_sizes_0, x = out_1_cast_fp16)[name = string("op_930_cast_fp16")];
            string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(974821440)))];
            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_930_cast_fp16_0)[name = string("query_states_1_cast_fp16")];
            string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(983210112)))];
            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_930_cast_fp16_0)[name = string("key_states_1_cast_fp16")];
            string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(984258752)))];
            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_930_cast_fp16_0)[name = string("value_states_1_cast_fp16")];
            tensor<int32, [4]> var_982 = const()[name = string("op_982"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_1_cast_fp16 = reshape(shape = var_982, x = query_states_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> var_986 = const()[name = string("op_986"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_987_cast_fp16 = reshape(shape = var_986, x = key_states_1_cast_fp16)[name = string("op_987_cast_fp16")];
            tensor<int32, [4]> var_993 = const()[name = string("op_993"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_994_cast_fp16 = reshape(shape = var_993, x = value_states_1_cast_fp16)[name = string("op_994_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_998_cast_fp16 = mul(x = x_1_cast_fp16, y = var_859_cast_fp16)[name = string("op_998_cast_fp16")];
            tensor<int32, [2]> var_999_split_sizes_0 = const()[name = string("op_999_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_999_axis_0 = const()[name = string("op_999_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_999_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_999_cast_fp16_1 = split(axis = var_999_axis_0, split_sizes = var_999_split_sizes_0, x = x_1_cast_fp16)[name = string("op_999_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_1001_cast_fp16 = mul(x = var_999_cast_fp16_1, y = const_7_promoted_to_fp16)[name = string("op_1001_cast_fp16")];
            int32 var_1003 = const()[name = string("op_1003"), val = int32(-2)];
            bool var_1004_interleave_0 = const()[name = string("op_1004_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1004_cast_fp16 = concat(axis = var_1003, interleave = var_1004_interleave_0, values = (var_1001_cast_fp16, var_999_cast_fp16_0))[name = string("op_1004_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1005_cast_fp16 = mul(x = var_1004_cast_fp16, y = var_866_cast_fp16)[name = string("op_1005_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_3_cast_fp16 = add(x = var_998_cast_fp16, y = var_1005_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1011_cast_fp16 = mul(x = var_987_cast_fp16, y = var_859_cast_fp16)[name = string("op_1011_cast_fp16")];
            tensor<int32, [2]> var_1012_split_sizes_0 = const()[name = string("op_1012_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1012_axis_0 = const()[name = string("op_1012_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1012_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1012_cast_fp16_1 = split(axis = var_1012_axis_0, split_sizes = var_1012_split_sizes_0, x = var_987_cast_fp16)[name = string("op_1012_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_1014_cast_fp16 = mul(x = var_1012_cast_fp16_1, y = const_8_promoted_to_fp16)[name = string("op_1014_cast_fp16")];
            int32 var_1016 = const()[name = string("op_1016"), val = int32(-2)];
            bool var_1017_interleave_0 = const()[name = string("op_1017_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1017_cast_fp16 = concat(axis = var_1016, interleave = var_1017_interleave_0, values = (var_1014_cast_fp16, var_1012_cast_fp16_0))[name = string("op_1017_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1018_cast_fp16 = mul(x = var_1017_cast_fp16, y = var_866_cast_fp16)[name = string("op_1018_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_5_cast_fp16 = add(x = var_1011_cast_fp16, y = var_1018_cast_fp16)[name = string("key_states_5_cast_fp16")];
            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<fp16, [28, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor<int32, [1]>([1])];
            int32 concat_3_axis_0 = const()[name = string("concat_3_axis_0"), val = int32(0)];
            bool concat_3_interleave_0 = const()[name = string("concat_3_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_3 = concat(axis = concat_3_axis_0, interleave = concat_3_interleave_0, values = (expand_dims_1, expand_dims_2, position_id, expand_dims_4))[name = string("concat_3")];
            tensor<int32, [1]> concat_4_values1_0 = const()[name = string("concat_4_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_4_values3_0 = const()[name = string("concat_4_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_4_axis_0 = const()[name = string("concat_4_axis_0"), val = int32(0)];
            bool concat_4_interleave_0 = const()[name = string("concat_4_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_4 = concat(axis = concat_4_axis_0, interleave = concat_4_interleave_0, values = (expand_dims_5, concat_4_values1_0, cache_position_end, concat_4_values3_0))[name = string("concat_4")];
            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_82")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_3, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_4, 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, [28, 2, 2048, 128]> coreml_update_state_56 = read_state(input = key_cache)[name = string("coreml_update_state_56")];
            tensor<fp16, [28, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            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_994_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_3, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_4, 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, [28, 2, 2048, 128]> coreml_update_state_57 = read_state(input = value_cache)[name = string("coreml_update_state_57")];
            tensor<int32, [4]> var_1088_begin_0 = const()[name = string("op_1088_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1088_end_0 = const()[name = string("op_1088_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_1088_end_mask_0 = const()[name = string("op_1088_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1088_cast_fp16 = slice_by_index(begin = var_1088_begin_0, end = var_1088_end_0, end_mask = var_1088_end_mask_0, x = coreml_update_state_56)[name = string("op_1088_cast_fp16")];
            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1091_axis_0 = const()[name = string("op_1091_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1091_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1091_cast_fp16_1 = split(axis = var_1091_axis_0, split_sizes = tile_0, x = var_1088_cast_fp16)[name = string("op_1091_cast_fp16")];
            tensor<int32, [4]> var_1098_begin_0 = const()[name = string("op_1098_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1098_end_0 = const()[name = string("op_1098_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_1098_end_mask_0 = const()[name = string("op_1098_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1098_cast_fp16 = slice_by_index(begin = var_1098_begin_0, end = var_1098_end_0, end_mask = var_1098_end_mask_0, x = coreml_update_state_57)[name = string("op_1098_cast_fp16")];
            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1101_axis_0 = const()[name = string("op_1101_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1101_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1101_cast_fp16_1 = split(axis = var_1101_axis_0, split_sizes = tile_1, x = var_1098_cast_fp16)[name = string("op_1101_cast_fp16")];
            tensor<int32, [2]> var_1104_split_sizes_0 = const()[name = string("op_1104_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1104_axis_0 = const()[name = string("op_1104_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1104_0, tensor<fp16, [1, 8, 128, 8]> var_1104_1 = split(axis = var_1104_axis_0, split_sizes = var_1104_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_1104")];
            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_1091_cast_fp16_0, y = var_1104_0)[name = string("attn_weights_1_cast_fp16")];
            fp16 var_1107_to_fp16 = const()[name = string("op_1107_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_1107_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_1111 = const()[name = string("op_1111"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_7_cast_fp16 = softmax(axis = var_1111, x = attn_weights_5_cast_fp16)[name = string("attn_weights_7_cast_fp16")];
            bool var_1117_transpose_x_1 = const()[name = string("op_1117_transpose_x_1"), val = bool(true)];
            bool var_1117_transpose_y_1 = const()[name = string("op_1117_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1117_cast_fp16 = matmul(transpose_x = var_1117_transpose_x_1, transpose_y = var_1117_transpose_y_1, x = attn_weights_7_cast_fp16, y = var_1101_cast_fp16_0)[name = string("op_1117_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_1091_cast_fp16_1, y = var_1104_1)[name = string("attn_weights_9_cast_fp16")];
            fp16 var_1119_to_fp16 = const()[name = string("op_1119_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_1119_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_1123 = const()[name = string("op_1123"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_15_cast_fp16 = softmax(axis = var_1123, 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_1101_cast_fp16_1)[name = string("attn_output_1_cast_fp16")];
            int32 var_1131 = const()[name = string("op_1131"), 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_1131, interleave = attn_output_3_interleave_0, values = (var_1117_cast_fp16, attn_output_1_cast_fp16))[name = string("attn_output_3_cast_fp16")];
            tensor<int32, [4]> var_1135_perm_0 = const()[name = string("op_1135_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_1140 = const()[name = string("op_1140"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_1135_cast_fp16 = transpose(perm = var_1135_perm_0, x = attn_output_3_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_7_cast_fp16 = reshape(shape = var_1140, x = var_1135_cast_fp16)[name = string("attn_output_7_cast_fp16")];
            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(985307392)))];
            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")];
            int32 var_1166 = const()[name = string("op_1166"), val = int32(1)];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1168_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_1168_cast_fp16")];
            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_1166, interleave = doubled_5_interleave_0, values = (hidden_states_5_cast_fp16, var_1168_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(993696064)))];
            fp16 var_1178_to_fp16 = const()[name = string("op_1178_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_1178_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [2]> var_1189_split_sizes_0 = const()[name = string("op_1189_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1189_axis_0 = const()[name = string("op_1189_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1189_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1189_cast_fp16_1 = split(axis = var_1189_axis_0, split_sizes = var_1189_split_sizes_0, x = out_3_cast_fp16)[name = string("op_1189_cast_fp16")];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(993704320)))];
            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_1189_cast_fp16_0)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1206_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_1206_cast_fp16")];
            string var_1212_pad_type_0 = const()[name = string("op_1212_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1212_strides_0 = const()[name = string("op_1212_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1212_pad_0 = const()[name = string("op_1212_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1212_dilations_0 = const()[name = string("op_1212_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1212_groups_0 = const()[name = string("op_1212_groups_0"), val = int32(1)];
            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(1018870208)))];
            tensor<fp16, [1, 6144, 1, 8]> var_1212_cast_fp16 = conv(dilations = var_1212_dilations_0, groups = var_1212_groups_0, pad = var_1212_pad_0, pad_type = var_1212_pad_type_0, strides = var_1212_strides_0, weight = layers_0_mlp_up_proj_weight_to_fp16, x = var_1189_cast_fp16_0)[name = string("op_1212_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_9_cast_fp16 = mul(x = var_1206_cast_fp16, y = var_1212_cast_fp16)[name = string("x_9_cast_fp16")];
            string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1044036096)))];
            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")];
            int32 var_1228 = const()[name = string("op_1228"), val = int32(1)];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1230_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_1230_cast_fp16")];
            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_1228, interleave = doubled_9_interleave_0, values = (hidden_states_9_cast_fp16, var_1230_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(1069201984)))];
            fp16 var_1240_to_fp16 = const()[name = string("op_1240_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_1240_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [2]> var_1251_split_sizes_0 = const()[name = string("op_1251_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1251_axis_0 = const()[name = string("op_1251_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1251_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1251_cast_fp16_1 = split(axis = var_1251_axis_0, split_sizes = var_1251_split_sizes_0, x = out_5_cast_fp16)[name = string("op_1251_cast_fp16")];
            string query_states_7_pad_type_0 = const()[name = string("query_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_7_strides_0 = const()[name = string("query_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1069210240)))];
            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_1251_cast_fp16_0)[name = string("query_states_7_cast_fp16")];
            string key_states_11_pad_type_0 = const()[name = string("key_states_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_11_strides_0 = const()[name = string("key_states_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1077598912)))];
            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_1251_cast_fp16_0)[name = string("key_states_11_cast_fp16")];
            string value_states_7_pad_type_0 = const()[name = string("value_states_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_7_strides_0 = const()[name = string("value_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_1251_cast_fp16_0)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> var_1303 = const()[name = string("op_1303"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_11_cast_fp16 = reshape(shape = var_1303, x = query_states_7_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> var_1307 = const()[name = string("op_1307"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_1308_cast_fp16 = reshape(shape = var_1307, x = key_states_11_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<int32, [4]> var_1314 = const()[name = string("op_1314"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_1315_cast_fp16 = reshape(shape = var_1314, x = value_states_7_cast_fp16)[name = string("op_1315_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_1319_cast_fp16 = mul(x = x_11_cast_fp16, y = var_859_cast_fp16)[name = string("op_1319_cast_fp16")];
            tensor<int32, [2]> var_1320_split_sizes_0 = const()[name = string("op_1320_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1320_axis_0 = const()[name = string("op_1320_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1320_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1320_cast_fp16_1 = split(axis = var_1320_axis_0, split_sizes = var_1320_split_sizes_0, x = x_11_cast_fp16)[name = string("op_1320_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_1322_cast_fp16 = mul(x = var_1320_cast_fp16_1, y = const_19_promoted_to_fp16)[name = string("op_1322_cast_fp16")];
            int32 var_1324 = const()[name = string("op_1324"), val = int32(-2)];
            bool var_1325_interleave_0 = const()[name = string("op_1325_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1325_cast_fp16 = concat(axis = var_1324, interleave = var_1325_interleave_0, values = (var_1322_cast_fp16, var_1320_cast_fp16_0))[name = string("op_1325_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1326_cast_fp16 = mul(x = var_1325_cast_fp16, y = var_866_cast_fp16)[name = string("op_1326_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_9_cast_fp16 = add(x = var_1319_cast_fp16, y = var_1326_cast_fp16)[name = string("query_states_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1332_cast_fp16 = mul(x = var_1308_cast_fp16, y = var_859_cast_fp16)[name = string("op_1332_cast_fp16")];
            tensor<int32, [2]> var_1333_split_sizes_0 = const()[name = string("op_1333_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1333_axis_0 = const()[name = string("op_1333_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1333_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1333_cast_fp16_1 = split(axis = var_1333_axis_0, split_sizes = var_1333_split_sizes_0, x = var_1308_cast_fp16)[name = string("op_1333_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_1335_cast_fp16 = mul(x = var_1333_cast_fp16_1, y = const_20_promoted_to_fp16)[name = string("op_1335_cast_fp16")];
            int32 var_1337 = const()[name = string("op_1337"), val = int32(-2)];
            bool var_1338_interleave_0 = const()[name = string("op_1338_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1338_cast_fp16 = concat(axis = var_1337, interleave = var_1338_interleave_0, values = (var_1335_cast_fp16, var_1333_cast_fp16_0))[name = string("op_1338_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1339_cast_fp16 = mul(x = var_1338_cast_fp16, y = var_866_cast_fp16)[name = string("op_1339_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_15_cast_fp16 = add(x = var_1332_cast_fp16, y = var_1339_cast_fp16)[name = string("key_states_15_cast_fp16")];
            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, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> expand_dims_14 = const()[name = string("expand_dims_14"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_17 = const()[name = string("expand_dims_17"), val = tensor<int32, [1]>([2])];
            int32 concat_11_axis_0 = const()[name = string("concat_11_axis_0"), val = int32(0)];
            bool concat_11_interleave_0 = const()[name = string("concat_11_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_11 = concat(axis = concat_11_axis_0, interleave = concat_11_interleave_0, values = (expand_dims_13, expand_dims_14, position_id, expand_dims_16))[name = string("concat_11")];
            tensor<int32, [1]> concat_12_values1_0 = const()[name = string("concat_12_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_12_values3_0 = const()[name = string("concat_12_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_12_axis_0 = const()[name = string("concat_12_axis_0"), val = int32(0)];
            bool concat_12_interleave_0 = const()[name = string("concat_12_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_12 = concat(axis = concat_12_axis_0, interleave = concat_12_interleave_0, values = (expand_dims_17, concat_12_values1_0, cache_position_end, concat_12_values3_0))[name = string("concat_12")];
            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_79")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_11, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_12, 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, [28, 2, 2048, 128]> coreml_update_state_58 = read_state(input = key_cache)[name = string("coreml_update_state_58")];
            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_1315_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_11, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_12, 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, [28, 2, 2048, 128]> coreml_update_state_59 = read_state(input = value_cache)[name = string("coreml_update_state_59")];
            tensor<int32, [4]> var_1409_begin_0 = const()[name = string("op_1409_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1409_end_0 = const()[name = string("op_1409_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1409_end_mask_0 = const()[name = string("op_1409_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1409_cast_fp16 = slice_by_index(begin = var_1409_begin_0, end = var_1409_end_0, end_mask = var_1409_end_mask_0, x = coreml_update_state_58)[name = string("op_1409_cast_fp16")];
            tensor<int32, [2]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1412_axis_0 = const()[name = string("op_1412_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1412_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1412_cast_fp16_1 = split(axis = var_1412_axis_0, split_sizes = tile_2, x = var_1409_cast_fp16)[name = string("op_1412_cast_fp16")];
            tensor<int32, [4]> var_1419_begin_0 = const()[name = string("op_1419_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1419_end_0 = const()[name = string("op_1419_end_0"), val = tensor<int32, [4]>([2, 2, 2048, 128])];
            tensor<bool, [4]> var_1419_end_mask_0 = const()[name = string("op_1419_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1419_cast_fp16 = slice_by_index(begin = var_1419_begin_0, end = var_1419_end_0, end_mask = var_1419_end_mask_0, x = coreml_update_state_59)[name = string("op_1419_cast_fp16")];
            tensor<int32, [2]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1422_axis_0 = const()[name = string("op_1422_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1422_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1422_cast_fp16_1 = split(axis = var_1422_axis_0, split_sizes = tile_3, x = var_1419_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<int32, [2]> var_1425_split_sizes_0 = const()[name = string("op_1425_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1425_axis_0 = const()[name = string("op_1425_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1425_0, tensor<fp16, [1, 8, 128, 8]> var_1425_1 = split(axis = var_1425_axis_0, split_sizes = var_1425_split_sizes_0, x = query_states_9_cast_fp16)[name = string("op_1425")];
            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_1412_cast_fp16_0, y = var_1425_0)[name = string("attn_weights_17_cast_fp16")];
            fp16 var_1428_to_fp16 = const()[name = string("op_1428_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_1428_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_1432 = const()[name = string("op_1432"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_23_cast_fp16 = softmax(axis = var_1432, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")];
            bool var_1438_transpose_x_1 = const()[name = string("op_1438_transpose_x_1"), val = bool(true)];
            bool var_1438_transpose_y_1 = const()[name = string("op_1438_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1438_cast_fp16 = matmul(transpose_x = var_1438_transpose_x_1, transpose_y = var_1438_transpose_y_1, x = attn_weights_23_cast_fp16, y = var_1422_cast_fp16_0)[name = string("op_1438_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_1412_cast_fp16_1, y = var_1425_1)[name = string("attn_weights_25_cast_fp16")];
            fp16 var_1440_to_fp16 = const()[name = string("op_1440_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_1440_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_1444 = const()[name = string("op_1444"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_31_cast_fp16 = softmax(axis = var_1444, 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_1422_cast_fp16_1)[name = string("attn_output_9_cast_fp16")];
            int32 var_1452 = const()[name = string("op_1452"), 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_1452, interleave = attn_output_11_interleave_0, values = (var_1438_cast_fp16, attn_output_9_cast_fp16))[name = string("attn_output_11_cast_fp16")];
            tensor<int32, [4]> var_1456_perm_0 = const()[name = string("op_1456_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_1461 = const()[name = string("op_1461"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_1456_cast_fp16 = transpose(perm = var_1456_perm_0, x = attn_output_11_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_15_cast_fp16 = reshape(shape = var_1461, x = var_1456_cast_fp16)[name = string("attn_output_15_cast_fp16")];
            string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1078647552)))];
            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")];
            int32 var_1487 = const()[name = string("op_1487"), val = int32(1)];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1489_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_1489_cast_fp16")];
            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_1487, interleave = doubled_13_interleave_0, values = (hidden_states_15_cast_fp16, var_1489_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(1087036224)))];
            fp16 var_1499_to_fp16 = const()[name = string("op_1499_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_1499_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [2]> var_1510_split_sizes_0 = const()[name = string("op_1510_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1510_axis_0 = const()[name = string("op_1510_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1510_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1510_cast_fp16_1 = split(axis = var_1510_axis_0, split_sizes = var_1510_split_sizes_0, x = out_7_cast_fp16)[name = string("op_1510_cast_fp16")];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1087044480)))];
            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_1510_cast_fp16_0)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1527_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_1527_cast_fp16")];
            string var_1533_pad_type_0 = const()[name = string("op_1533_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1533_strides_0 = const()[name = string("op_1533_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1533_pad_0 = const()[name = string("op_1533_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1533_dilations_0 = const()[name = string("op_1533_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1533_groups_0 = const()[name = string("op_1533_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1533_cast_fp16 = conv(dilations = var_1533_dilations_0, groups = var_1533_groups_0, pad = var_1533_pad_0, pad_type = var_1533_pad_type_0, strides = var_1533_strides_0, weight = layers_1_mlp_up_proj_weight_cast_fp16, x = var_1510_cast_fp16_0)[name = string("op_1533_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_19_cast_fp16 = mul(x = var_1527_cast_fp16, y = var_1533_cast_fp16)[name = string("x_19_cast_fp16")];
            string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1112210368)))];
            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")];
            int32 var_1549 = const()[name = string("op_1549"), val = int32(1)];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1551_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_1551_cast_fp16")];
            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_1549, interleave = doubled_17_interleave_0, values = (hidden_states_19_cast_fp16, var_1551_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(1137376256)))];
            fp16 var_1561_to_fp16 = const()[name = string("op_1561_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_1561_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [2]> var_1572_split_sizes_0 = const()[name = string("op_1572_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1572_axis_0 = const()[name = string("op_1572_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1572_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1572_cast_fp16_1 = split(axis = var_1572_axis_0, split_sizes = var_1572_split_sizes_0, x = out_9_cast_fp16)[name = string("op_1572_cast_fp16")];
            string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1137384512)))];
            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_1572_cast_fp16_0)[name = string("query_states_13_cast_fp16")];
            string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1145773184)))];
            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_1572_cast_fp16_0)[name = string("key_states_21_cast_fp16")];
            string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_1572_cast_fp16_0)[name = string("value_states_13_cast_fp16")];
            tensor<int32, [4]> var_1624 = const()[name = string("op_1624"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_21_cast_fp16 = reshape(shape = var_1624, x = query_states_13_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> var_1628 = const()[name = string("op_1628"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_1629_cast_fp16 = reshape(shape = var_1628, x = key_states_21_cast_fp16)[name = string("op_1629_cast_fp16")];
            tensor<int32, [4]> var_1635 = const()[name = string("op_1635"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_1636_cast_fp16 = reshape(shape = var_1635, x = value_states_13_cast_fp16)[name = string("op_1636_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_1640_cast_fp16 = mul(x = x_21_cast_fp16, y = var_859_cast_fp16)[name = string("op_1640_cast_fp16")];
            tensor<int32, [2]> var_1641_split_sizes_0 = const()[name = string("op_1641_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1641_axis_0 = const()[name = string("op_1641_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1641_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1641_cast_fp16_1 = split(axis = var_1641_axis_0, split_sizes = var_1641_split_sizes_0, x = x_21_cast_fp16)[name = string("op_1641_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_1643_cast_fp16 = mul(x = var_1641_cast_fp16_1, y = const_31_promoted_to_fp16)[name = string("op_1643_cast_fp16")];
            int32 var_1645 = const()[name = string("op_1645"), val = int32(-2)];
            bool var_1646_interleave_0 = const()[name = string("op_1646_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1646_cast_fp16 = concat(axis = var_1645, interleave = var_1646_interleave_0, values = (var_1643_cast_fp16, var_1641_cast_fp16_0))[name = string("op_1646_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1647_cast_fp16 = mul(x = var_1646_cast_fp16, y = var_866_cast_fp16)[name = string("op_1647_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_15_cast_fp16 = add(x = var_1640_cast_fp16, y = var_1647_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1653_cast_fp16 = mul(x = var_1629_cast_fp16, y = var_859_cast_fp16)[name = string("op_1653_cast_fp16")];
            tensor<int32, [2]> var_1654_split_sizes_0 = const()[name = string("op_1654_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1654_axis_0 = const()[name = string("op_1654_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1654_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1654_cast_fp16_1 = split(axis = var_1654_axis_0, split_sizes = var_1654_split_sizes_0, x = var_1629_cast_fp16)[name = string("op_1654_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_1656_cast_fp16 = mul(x = var_1654_cast_fp16_1, y = const_32_promoted_to_fp16)[name = string("op_1656_cast_fp16")];
            int32 var_1658 = const()[name = string("op_1658"), val = int32(-2)];
            bool var_1659_interleave_0 = const()[name = string("op_1659_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1659_cast_fp16 = concat(axis = var_1658, interleave = var_1659_interleave_0, values = (var_1656_cast_fp16, var_1654_cast_fp16_0))[name = string("op_1659_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1660_cast_fp16 = mul(x = var_1659_cast_fp16, y = var_866_cast_fp16)[name = string("op_1660_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_25_cast_fp16 = add(x = var_1653_cast_fp16, y = var_1660_cast_fp16)[name = string("key_states_25_cast_fp16")];
            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, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_26 = const()[name = string("expand_dims_26"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_28 = const()[name = string("expand_dims_28"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_29 = const()[name = string("expand_dims_29"), val = tensor<int32, [1]>([3])];
            int32 concat_19_axis_0 = const()[name = string("concat_19_axis_0"), val = int32(0)];
            bool concat_19_interleave_0 = const()[name = string("concat_19_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_19 = concat(axis = concat_19_axis_0, interleave = concat_19_interleave_0, values = (expand_dims_25, expand_dims_26, position_id, expand_dims_28))[name = string("concat_19")];
            tensor<int32, [1]> concat_20_values1_0 = const()[name = string("concat_20_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_20_values3_0 = const()[name = string("concat_20_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_20_axis_0 = const()[name = string("concat_20_axis_0"), val = int32(0)];
            bool concat_20_interleave_0 = const()[name = string("concat_20_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_20 = concat(axis = concat_20_axis_0, interleave = concat_20_interleave_0, values = (expand_dims_29, concat_20_values1_0, cache_position_end, concat_20_values3_0))[name = string("concat_20")];
            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_76")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_19, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_20, 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, [28, 2, 2048, 128]> coreml_update_state_60 = read_state(input = key_cache)[name = string("coreml_update_state_60")];
            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_1636_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_19, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_20, 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, [28, 2, 2048, 128]> coreml_update_state_61 = read_state(input = value_cache)[name = string("coreml_update_state_61")];
            tensor<int32, [4]> var_1730_begin_0 = const()[name = string("op_1730_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1730_end_0 = const()[name = string("op_1730_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1730_end_mask_0 = const()[name = string("op_1730_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1730_cast_fp16 = slice_by_index(begin = var_1730_begin_0, end = var_1730_end_0, end_mask = var_1730_end_mask_0, x = coreml_update_state_60)[name = string("op_1730_cast_fp16")];
            tensor<int32, [2]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1733_axis_0 = const()[name = string("op_1733_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1733_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1733_cast_fp16_1 = split(axis = var_1733_axis_0, split_sizes = tile_4, x = var_1730_cast_fp16)[name = string("op_1733_cast_fp16")];
            tensor<int32, [4]> var_1740_begin_0 = const()[name = string("op_1740_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1740_end_0 = const()[name = string("op_1740_end_0"), val = tensor<int32, [4]>([3, 2, 2048, 128])];
            tensor<bool, [4]> var_1740_end_mask_0 = const()[name = string("op_1740_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_1740_cast_fp16 = slice_by_index(begin = var_1740_begin_0, end = var_1740_end_0, end_mask = var_1740_end_mask_0, x = coreml_update_state_61)[name = string("op_1740_cast_fp16")];
            tensor<int32, [2]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1743_axis_0 = const()[name = string("op_1743_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_1743_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_1743_cast_fp16_1 = split(axis = var_1743_axis_0, split_sizes = tile_5, x = var_1740_cast_fp16)[name = string("op_1743_cast_fp16")];
            tensor<int32, [2]> var_1746_split_sizes_0 = const()[name = string("op_1746_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_1746_axis_0 = const()[name = string("op_1746_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_1746_0, tensor<fp16, [1, 8, 128, 8]> var_1746_1 = split(axis = var_1746_axis_0, split_sizes = var_1746_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_1746")];
            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_1733_cast_fp16_0, y = var_1746_0)[name = string("attn_weights_33_cast_fp16")];
            fp16 var_1749_to_fp16 = const()[name = string("op_1749_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_1749_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_1753 = const()[name = string("op_1753"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_39_cast_fp16 = softmax(axis = var_1753, x = attn_weights_37_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            bool var_1759_transpose_x_1 = const()[name = string("op_1759_transpose_x_1"), val = bool(true)];
            bool var_1759_transpose_y_1 = const()[name = string("op_1759_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_1759_cast_fp16 = matmul(transpose_x = var_1759_transpose_x_1, transpose_y = var_1759_transpose_y_1, x = attn_weights_39_cast_fp16, y = var_1743_cast_fp16_0)[name = string("op_1759_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_1733_cast_fp16_1, y = var_1746_1)[name = string("attn_weights_41_cast_fp16")];
            fp16 var_1761_to_fp16 = const()[name = string("op_1761_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_1761_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_1765 = const()[name = string("op_1765"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_47_cast_fp16 = softmax(axis = var_1765, 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_1743_cast_fp16_1)[name = string("attn_output_17_cast_fp16")];
            int32 var_1773 = const()[name = string("op_1773"), 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_1773, interleave = attn_output_19_interleave_0, values = (var_1759_cast_fp16, attn_output_17_cast_fp16))[name = string("attn_output_19_cast_fp16")];
            tensor<int32, [4]> var_1777_perm_0 = const()[name = string("op_1777_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_1782 = const()[name = string("op_1782"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_1777_cast_fp16 = transpose(perm = var_1777_perm_0, x = attn_output_19_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_23_cast_fp16 = reshape(shape = var_1782, x = var_1777_cast_fp16)[name = string("attn_output_23_cast_fp16")];
            string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1146821824)))];
            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")];
            int32 var_1808 = const()[name = string("op_1808"), val = int32(1)];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1810_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_1810_cast_fp16")];
            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_1808, interleave = doubled_21_interleave_0, values = (hidden_states_25_cast_fp16, var_1810_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(1155210496)))];
            fp16 var_1820_to_fp16 = const()[name = string("op_1820_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_1820_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [2]> var_1831_split_sizes_0 = const()[name = string("op_1831_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1831_axis_0 = const()[name = string("op_1831_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1831_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1831_cast_fp16_1 = split(axis = var_1831_axis_0, split_sizes = var_1831_split_sizes_0, x = out_11_cast_fp16)[name = string("op_1831_cast_fp16")];
            string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1155218752)))];
            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_1831_cast_fp16_0)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_1848_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_1848_cast_fp16")];
            string var_1854_pad_type_0 = const()[name = string("op_1854_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1854_strides_0 = const()[name = string("op_1854_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1854_pad_0 = const()[name = string("op_1854_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1854_dilations_0 = const()[name = string("op_1854_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1854_groups_0 = const()[name = string("op_1854_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_1854_cast_fp16 = conv(dilations = var_1854_dilations_0, groups = var_1854_groups_0, pad = var_1854_pad_0, pad_type = var_1854_pad_type_0, strides = var_1854_strides_0, weight = layers_2_mlp_up_proj_weight_cast_fp16, x = var_1831_cast_fp16_0)[name = string("op_1854_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_29_cast_fp16 = mul(x = var_1848_cast_fp16, y = var_1854_cast_fp16)[name = string("x_29_cast_fp16")];
            string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1180384640)))];
            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")];
            int32 var_1870 = const()[name = string("op_1870"), val = int32(1)];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_1872_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_1872_cast_fp16")];
            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_1870, interleave = doubled_25_interleave_0, values = (hidden_states_29_cast_fp16, var_1872_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(1205550528)))];
            fp16 var_1882_to_fp16 = const()[name = string("op_1882_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_1882_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [2]> var_1893_split_sizes_0 = const()[name = string("op_1893_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_1893_axis_0 = const()[name = string("op_1893_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_1893_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_1893_cast_fp16_1 = split(axis = var_1893_axis_0, split_sizes = var_1893_split_sizes_0, x = out_13_cast_fp16)[name = string("op_1893_cast_fp16")];
            string query_states_19_pad_type_0 = const()[name = string("query_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_19_strides_0 = const()[name = string("query_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1205558784)))];
            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_1893_cast_fp16_0)[name = string("query_states_19_cast_fp16")];
            string key_states_31_pad_type_0 = const()[name = string("key_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_31_strides_0 = const()[name = string("key_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1213947456)))];
            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_1893_cast_fp16_0)[name = string("key_states_31_cast_fp16")];
            string value_states_19_pad_type_0 = const()[name = string("value_states_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_19_strides_0 = const()[name = string("value_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_1893_cast_fp16_0)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> var_1945 = const()[name = string("op_1945"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_31_cast_fp16 = reshape(shape = var_1945, x = query_states_19_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> var_1949 = const()[name = string("op_1949"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_1950_cast_fp16 = reshape(shape = var_1949, x = key_states_31_cast_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [4]> var_1956 = const()[name = string("op_1956"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_1957_cast_fp16 = reshape(shape = var_1956, x = value_states_19_cast_fp16)[name = string("op_1957_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_1961_cast_fp16 = mul(x = x_31_cast_fp16, y = var_859_cast_fp16)[name = string("op_1961_cast_fp16")];
            tensor<int32, [2]> var_1962_split_sizes_0 = const()[name = string("op_1962_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1962_axis_0 = const()[name = string("op_1962_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_1962_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_1962_cast_fp16_1 = split(axis = var_1962_axis_0, split_sizes = var_1962_split_sizes_0, x = x_31_cast_fp16)[name = string("op_1962_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_1964_cast_fp16 = mul(x = var_1962_cast_fp16_1, y = const_43_promoted_to_fp16)[name = string("op_1964_cast_fp16")];
            int32 var_1966 = const()[name = string("op_1966"), val = int32(-2)];
            bool var_1967_interleave_0 = const()[name = string("op_1967_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_1967_cast_fp16 = concat(axis = var_1966, interleave = var_1967_interleave_0, values = (var_1964_cast_fp16, var_1962_cast_fp16_0))[name = string("op_1967_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_1968_cast_fp16 = mul(x = var_1967_cast_fp16, y = var_866_cast_fp16)[name = string("op_1968_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_21_cast_fp16 = add(x = var_1961_cast_fp16, y = var_1968_cast_fp16)[name = string("query_states_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1974_cast_fp16 = mul(x = var_1950_cast_fp16, y = var_859_cast_fp16)[name = string("op_1974_cast_fp16")];
            tensor<int32, [2]> var_1975_split_sizes_0 = const()[name = string("op_1975_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1975_axis_0 = const()[name = string("op_1975_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_1975_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_1975_cast_fp16_1 = split(axis = var_1975_axis_0, split_sizes = var_1975_split_sizes_0, x = var_1950_cast_fp16)[name = string("op_1975_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_1977_cast_fp16 = mul(x = var_1975_cast_fp16_1, y = const_44_promoted_to_fp16)[name = string("op_1977_cast_fp16")];
            int32 var_1979 = const()[name = string("op_1979"), val = int32(-2)];
            bool var_1980_interleave_0 = const()[name = string("op_1980_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_1980_cast_fp16 = concat(axis = var_1979, interleave = var_1980_interleave_0, values = (var_1977_cast_fp16, var_1975_cast_fp16_0))[name = string("op_1980_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_1981_cast_fp16 = mul(x = var_1980_cast_fp16, y = var_866_cast_fp16)[name = string("op_1981_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_35_cast_fp16 = add(x = var_1974_cast_fp16, y = var_1981_cast_fp16)[name = string("key_states_35_cast_fp16")];
            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, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([3])];
            tensor<int32, [1]> expand_dims_38 = const()[name = string("expand_dims_38"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_40 = const()[name = string("expand_dims_40"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_41 = const()[name = string("expand_dims_41"), val = tensor<int32, [1]>([4])];
            int32 concat_27_axis_0 = const()[name = string("concat_27_axis_0"), val = int32(0)];
            bool concat_27_interleave_0 = const()[name = string("concat_27_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_27 = concat(axis = concat_27_axis_0, interleave = concat_27_interleave_0, values = (expand_dims_37, expand_dims_38, position_id, expand_dims_40))[name = string("concat_27")];
            tensor<int32, [1]> concat_28_values1_0 = const()[name = string("concat_28_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_28_values3_0 = const()[name = string("concat_28_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_28_axis_0 = const()[name = string("concat_28_axis_0"), val = int32(0)];
            bool concat_28_interleave_0 = const()[name = string("concat_28_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_28 = concat(axis = concat_28_axis_0, interleave = concat_28_interleave_0, values = (expand_dims_41, concat_28_values1_0, cache_position_end, concat_28_values3_0))[name = string("concat_28")];
            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_73")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_27, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_28, 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, [28, 2, 2048, 128]> coreml_update_state_62 = read_state(input = key_cache)[name = string("coreml_update_state_62")];
            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_1957_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_27, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_28, 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, [28, 2, 2048, 128]> coreml_update_state_63 = read_state(input = value_cache)[name = string("coreml_update_state_63")];
            tensor<int32, [4]> var_2051_begin_0 = const()[name = string("op_2051_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2051_end_0 = const()[name = string("op_2051_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_2051_end_mask_0 = const()[name = string("op_2051_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2051_cast_fp16 = slice_by_index(begin = var_2051_begin_0, end = var_2051_end_0, end_mask = var_2051_end_mask_0, x = coreml_update_state_62)[name = string("op_2051_cast_fp16")];
            tensor<int32, [2]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2054_axis_0 = const()[name = string("op_2054_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2054_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2054_cast_fp16_1 = split(axis = var_2054_axis_0, split_sizes = tile_6, x = var_2051_cast_fp16)[name = string("op_2054_cast_fp16")];
            tensor<int32, [4]> var_2061_begin_0 = const()[name = string("op_2061_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2061_end_0 = const()[name = string("op_2061_end_0"), val = tensor<int32, [4]>([4, 2, 2048, 128])];
            tensor<bool, [4]> var_2061_end_mask_0 = const()[name = string("op_2061_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2061_cast_fp16 = slice_by_index(begin = var_2061_begin_0, end = var_2061_end_0, end_mask = var_2061_end_mask_0, x = coreml_update_state_63)[name = string("op_2061_cast_fp16")];
            tensor<int32, [2]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2064_axis_0 = const()[name = string("op_2064_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2064_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2064_cast_fp16_1 = split(axis = var_2064_axis_0, split_sizes = tile_7, x = var_2061_cast_fp16)[name = string("op_2064_cast_fp16")];
            tensor<int32, [2]> var_2067_split_sizes_0 = const()[name = string("op_2067_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2067_axis_0 = const()[name = string("op_2067_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2067_0, tensor<fp16, [1, 8, 128, 8]> var_2067_1 = split(axis = var_2067_axis_0, split_sizes = var_2067_split_sizes_0, x = query_states_21_cast_fp16)[name = string("op_2067")];
            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_2054_cast_fp16_0, y = var_2067_0)[name = string("attn_weights_49_cast_fp16")];
            fp16 var_2070_to_fp16 = const()[name = string("op_2070_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_2070_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_2074 = const()[name = string("op_2074"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_55_cast_fp16 = softmax(axis = var_2074, x = attn_weights_53_cast_fp16)[name = string("attn_weights_55_cast_fp16")];
            bool var_2080_transpose_x_1 = const()[name = string("op_2080_transpose_x_1"), val = bool(true)];
            bool var_2080_transpose_y_1 = const()[name = string("op_2080_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2080_cast_fp16 = matmul(transpose_x = var_2080_transpose_x_1, transpose_y = var_2080_transpose_y_1, x = attn_weights_55_cast_fp16, y = var_2064_cast_fp16_0)[name = string("op_2080_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_2054_cast_fp16_1, y = var_2067_1)[name = string("attn_weights_57_cast_fp16")];
            fp16 var_2082_to_fp16 = const()[name = string("op_2082_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_2082_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_2086 = const()[name = string("op_2086"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_63_cast_fp16 = softmax(axis = var_2086, 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_2064_cast_fp16_1)[name = string("attn_output_25_cast_fp16")];
            int32 var_2094 = const()[name = string("op_2094"), 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_2094, interleave = attn_output_27_interleave_0, values = (var_2080_cast_fp16, attn_output_25_cast_fp16))[name = string("attn_output_27_cast_fp16")];
            tensor<int32, [4]> var_2098_perm_0 = const()[name = string("op_2098_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_2103 = const()[name = string("op_2103"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_2098_cast_fp16 = transpose(perm = var_2098_perm_0, x = attn_output_27_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_31_cast_fp16 = reshape(shape = var_2103, x = var_2098_cast_fp16)[name = string("attn_output_31_cast_fp16")];
            string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_2129 = const()[name = string("op_2129"), val = int32(1)];
            fp16 const_49_promoted_to_fp16 = const()[name = string("const_49_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2131_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = const_49_promoted_to_fp16)[name = string("op_2131_cast_fp16")];
            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_2129, interleave = doubled_29_interleave_0, values = (hidden_states_35_cast_fp16, var_2131_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(1214996096)))];
            fp16 var_2141_to_fp16 = const()[name = string("op_2141_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_2141_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [2]> var_2152_split_sizes_0 = const()[name = string("op_2152_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2152_axis_0 = const()[name = string("op_2152_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2152_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2152_cast_fp16_1 = split(axis = var_2152_axis_0, split_sizes = var_2152_split_sizes_0, x = out_15_cast_fp16)[name = string("op_2152_cast_fp16")];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1215004352)))];
            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_2152_cast_fp16_0)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2169_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_2169_cast_fp16")];
            string var_2175_pad_type_0 = const()[name = string("op_2175_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2175_strides_0 = const()[name = string("op_2175_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2175_pad_0 = const()[name = string("op_2175_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2175_dilations_0 = const()[name = string("op_2175_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2175_groups_0 = const()[name = string("op_2175_groups_0"), val = int32(1)];
            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(1240170240)))];
            tensor<fp16, [1, 6144, 1, 8]> var_2175_cast_fp16 = conv(dilations = var_2175_dilations_0, groups = var_2175_groups_0, pad = var_2175_pad_0, pad_type = var_2175_pad_type_0, strides = var_2175_strides_0, weight = layers_3_mlp_up_proj_weight_to_fp16, x = var_2152_cast_fp16_0)[name = string("op_2175_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_39_cast_fp16 = mul(x = var_2169_cast_fp16, y = var_2175_cast_fp16)[name = string("x_39_cast_fp16")];
            string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_2191 = const()[name = string("op_2191"), val = int32(1)];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2193_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_2193_cast_fp16")];
            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_2191, interleave = doubled_33_interleave_0, values = (hidden_states_39_cast_fp16, var_2193_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(1265336128)))];
            fp16 var_2203_to_fp16 = const()[name = string("op_2203_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_2203_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [2]> var_2214_split_sizes_0 = const()[name = string("op_2214_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2214_axis_0 = const()[name = string("op_2214_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2214_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2214_cast_fp16_1 = split(axis = var_2214_axis_0, split_sizes = var_2214_split_sizes_0, x = out_17_cast_fp16)[name = string("op_2214_cast_fp16")];
            string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1265344384)))];
            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_2214_cast_fp16_0)[name = string("query_states_25_cast_fp16")];
            string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1273733056)))];
            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_2214_cast_fp16_0)[name = string("key_states_41_cast_fp16")];
            string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_2214_cast_fp16_0)[name = string("value_states_25_cast_fp16")];
            tensor<int32, [4]> var_2266 = const()[name = string("op_2266"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_41_cast_fp16 = reshape(shape = var_2266, x = query_states_25_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> var_2270 = const()[name = string("op_2270"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_2271_cast_fp16 = reshape(shape = var_2270, x = key_states_41_cast_fp16)[name = string("op_2271_cast_fp16")];
            tensor<int32, [4]> var_2277 = const()[name = string("op_2277"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_2278_cast_fp16 = reshape(shape = var_2277, x = value_states_25_cast_fp16)[name = string("op_2278_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_2282_cast_fp16 = mul(x = x_41_cast_fp16, y = var_859_cast_fp16)[name = string("op_2282_cast_fp16")];
            tensor<int32, [2]> var_2283_split_sizes_0 = const()[name = string("op_2283_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2283_axis_0 = const()[name = string("op_2283_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2283_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2283_cast_fp16_1 = split(axis = var_2283_axis_0, split_sizes = var_2283_split_sizes_0, x = x_41_cast_fp16)[name = string("op_2283_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_2285_cast_fp16 = mul(x = var_2283_cast_fp16_1, y = const_55_promoted_to_fp16)[name = string("op_2285_cast_fp16")];
            int32 var_2287 = const()[name = string("op_2287"), val = int32(-2)];
            bool var_2288_interleave_0 = const()[name = string("op_2288_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2288_cast_fp16 = concat(axis = var_2287, interleave = var_2288_interleave_0, values = (var_2285_cast_fp16, var_2283_cast_fp16_0))[name = string("op_2288_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2289_cast_fp16 = mul(x = var_2288_cast_fp16, y = var_866_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_27_cast_fp16 = add(x = var_2282_cast_fp16, y = var_2289_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2295_cast_fp16 = mul(x = var_2271_cast_fp16, y = var_859_cast_fp16)[name = string("op_2295_cast_fp16")];
            tensor<int32, [2]> var_2296_split_sizes_0 = const()[name = string("op_2296_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2296_axis_0 = const()[name = string("op_2296_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2296_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2296_cast_fp16_1 = split(axis = var_2296_axis_0, split_sizes = var_2296_split_sizes_0, x = var_2271_cast_fp16)[name = string("op_2296_cast_fp16")];
            fp16 const_56_promoted_to_fp16 = const()[name = string("const_56_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_2298_cast_fp16 = mul(x = var_2296_cast_fp16_1, y = const_56_promoted_to_fp16)[name = string("op_2298_cast_fp16")];
            int32 var_2300 = const()[name = string("op_2300"), val = int32(-2)];
            bool var_2301_interleave_0 = const()[name = string("op_2301_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2301_cast_fp16 = concat(axis = var_2300, interleave = var_2301_interleave_0, values = (var_2298_cast_fp16, var_2296_cast_fp16_0))[name = string("op_2301_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2302_cast_fp16 = mul(x = var_2301_cast_fp16, y = var_866_cast_fp16)[name = string("op_2302_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_45_cast_fp16 = add(x = var_2295_cast_fp16, y = var_2302_cast_fp16)[name = string("key_states_45_cast_fp16")];
            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, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([4])];
            tensor<int32, [1]> expand_dims_50 = const()[name = string("expand_dims_50"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_52 = const()[name = string("expand_dims_52"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_53 = const()[name = string("expand_dims_53"), val = tensor<int32, [1]>([5])];
            int32 concat_35_axis_0 = const()[name = string("concat_35_axis_0"), val = int32(0)];
            bool concat_35_interleave_0 = const()[name = string("concat_35_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_35 = concat(axis = concat_35_axis_0, interleave = concat_35_interleave_0, values = (expand_dims_49, expand_dims_50, position_id, expand_dims_52))[name = string("concat_35")];
            tensor<int32, [1]> concat_36_values1_0 = const()[name = string("concat_36_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_36_values3_0 = const()[name = string("concat_36_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_36_axis_0 = const()[name = string("concat_36_axis_0"), val = int32(0)];
            bool concat_36_interleave_0 = const()[name = string("concat_36_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_36 = concat(axis = concat_36_axis_0, interleave = concat_36_interleave_0, values = (expand_dims_53, concat_36_values1_0, cache_position_end, concat_36_values3_0))[name = string("concat_36")];
            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_70")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_35, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_36, 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, [28, 2, 2048, 128]> coreml_update_state_64 = read_state(input = key_cache)[name = string("coreml_update_state_64")];
            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_2278_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_35, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_36, 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, [28, 2, 2048, 128]> coreml_update_state_65 = read_state(input = value_cache)[name = string("coreml_update_state_65")];
            tensor<int32, [4]> var_2372_begin_0 = const()[name = string("op_2372_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_2372_end_0 = const()[name = string("op_2372_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_2372_end_mask_0 = const()[name = string("op_2372_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2372_cast_fp16 = slice_by_index(begin = var_2372_begin_0, end = var_2372_end_0, end_mask = var_2372_end_mask_0, x = coreml_update_state_64)[name = string("op_2372_cast_fp16")];
            tensor<int32, [2]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2375_axis_0 = const()[name = string("op_2375_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2375_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2375_cast_fp16_1 = split(axis = var_2375_axis_0, split_sizes = tile_8, x = var_2372_cast_fp16)[name = string("op_2375_cast_fp16")];
            tensor<int32, [4]> var_2382_begin_0 = const()[name = string("op_2382_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_2382_end_0 = const()[name = string("op_2382_end_0"), val = tensor<int32, [4]>([5, 2, 2048, 128])];
            tensor<bool, [4]> var_2382_end_mask_0 = const()[name = string("op_2382_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2382_cast_fp16 = slice_by_index(begin = var_2382_begin_0, end = var_2382_end_0, end_mask = var_2382_end_mask_0, x = coreml_update_state_65)[name = string("op_2382_cast_fp16")];
            tensor<int32, [2]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2385_axis_0 = const()[name = string("op_2385_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2385_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2385_cast_fp16_1 = split(axis = var_2385_axis_0, split_sizes = tile_9, x = var_2382_cast_fp16)[name = string("op_2385_cast_fp16")];
            tensor<int32, [2]> var_2388_split_sizes_0 = const()[name = string("op_2388_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2388_axis_0 = const()[name = string("op_2388_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2388_0, tensor<fp16, [1, 8, 128, 8]> var_2388_1 = split(axis = var_2388_axis_0, split_sizes = var_2388_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_2388")];
            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_2375_cast_fp16_0, y = var_2388_0)[name = string("attn_weights_65_cast_fp16")];
            fp16 var_2391_to_fp16 = const()[name = string("op_2391_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_2391_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_2395 = const()[name = string("op_2395"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_71_cast_fp16 = softmax(axis = var_2395, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")];
            bool var_2401_transpose_x_1 = const()[name = string("op_2401_transpose_x_1"), val = bool(true)];
            bool var_2401_transpose_y_1 = const()[name = string("op_2401_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2401_cast_fp16 = matmul(transpose_x = var_2401_transpose_x_1, transpose_y = var_2401_transpose_y_1, x = attn_weights_71_cast_fp16, y = var_2385_cast_fp16_0)[name = string("op_2401_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_2375_cast_fp16_1, y = var_2388_1)[name = string("attn_weights_73_cast_fp16")];
            fp16 var_2403_to_fp16 = const()[name = string("op_2403_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_2403_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_2407 = const()[name = string("op_2407"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_79_cast_fp16 = softmax(axis = var_2407, 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_2385_cast_fp16_1)[name = string("attn_output_33_cast_fp16")];
            int32 var_2415 = const()[name = string("op_2415"), 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_2415, interleave = attn_output_35_interleave_0, values = (var_2401_cast_fp16, attn_output_33_cast_fp16))[name = string("attn_output_35_cast_fp16")];
            tensor<int32, [4]> var_2419_perm_0 = const()[name = string("op_2419_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_2424 = const()[name = string("op_2424"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_2419_cast_fp16 = transpose(perm = var_2419_perm_0, x = attn_output_35_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_39_cast_fp16 = reshape(shape = var_2424, x = var_2419_cast_fp16)[name = string("attn_output_39_cast_fp16")];
            string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_2450 = const()[name = string("op_2450"), val = int32(1)];
            fp16 const_61_promoted_to_fp16 = const()[name = string("const_61_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2452_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = const_61_promoted_to_fp16)[name = string("op_2452_cast_fp16")];
            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_2450, interleave = doubled_37_interleave_0, values = (hidden_states_45_cast_fp16, var_2452_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(1274781696)))];
            fp16 var_2462_to_fp16 = const()[name = string("op_2462_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_2462_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [2]> var_2473_split_sizes_0 = const()[name = string("op_2473_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2473_axis_0 = const()[name = string("op_2473_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2473_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2473_cast_fp16_1 = split(axis = var_2473_axis_0, split_sizes = var_2473_split_sizes_0, x = out_19_cast_fp16)[name = string("op_2473_cast_fp16")];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_2473_cast_fp16_0)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2490_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_2490_cast_fp16")];
            string var_2496_pad_type_0 = const()[name = string("op_2496_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2496_strides_0 = const()[name = string("op_2496_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2496_pad_0 = const()[name = string("op_2496_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2496_dilations_0 = const()[name = string("op_2496_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2496_groups_0 = const()[name = string("op_2496_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2496_cast_fp16 = conv(dilations = var_2496_dilations_0, groups = var_2496_groups_0, pad = var_2496_pad_0, pad_type = var_2496_pad_type_0, strides = var_2496_strides_0, weight = layers_4_mlp_up_proj_weight_cast_fp16, x = var_2473_cast_fp16_0)[name = string("op_2496_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_49_cast_fp16 = mul(x = var_2490_cast_fp16, y = var_2496_cast_fp16)[name = string("x_49_cast_fp16")];
            string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_2512 = const()[name = string("op_2512"), val = int32(1)];
            fp16 const_63_promoted_to_fp16 = const()[name = string("const_63_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2514_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = const_63_promoted_to_fp16)[name = string("op_2514_cast_fp16")];
            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_2512, interleave = doubled_41_interleave_0, values = (hidden_states_49_cast_fp16, var_2514_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(1274789952)))];
            fp16 var_2524_to_fp16 = const()[name = string("op_2524_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_2524_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [2]> var_2535_split_sizes_0 = const()[name = string("op_2535_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2535_axis_0 = const()[name = string("op_2535_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2535_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2535_cast_fp16_1 = split(axis = var_2535_axis_0, split_sizes = var_2535_split_sizes_0, x = out_21_cast_fp16)[name = string("op_2535_cast_fp16")];
            string query_states_31_pad_type_0 = const()[name = string("query_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_31_strides_0 = const()[name = string("query_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1274798208)))];
            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_2535_cast_fp16_0)[name = string("query_states_31_cast_fp16")];
            string key_states_51_pad_type_0 = const()[name = string("key_states_51_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_51_strides_0 = const()[name = string("key_states_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1283186880)))];
            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_2535_cast_fp16_0)[name = string("key_states_51_cast_fp16")];
            string value_states_31_pad_type_0 = const()[name = string("value_states_31_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_31_strides_0 = const()[name = string("value_states_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_2535_cast_fp16_0)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> var_2587 = const()[name = string("op_2587"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_51_cast_fp16 = reshape(shape = var_2587, x = query_states_31_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> var_2591 = const()[name = string("op_2591"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_2592_cast_fp16 = reshape(shape = var_2591, x = key_states_51_cast_fp16)[name = string("op_2592_cast_fp16")];
            tensor<int32, [4]> var_2598 = const()[name = string("op_2598"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_2599_cast_fp16 = reshape(shape = var_2598, x = value_states_31_cast_fp16)[name = string("op_2599_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_2603_cast_fp16 = mul(x = x_51_cast_fp16, y = var_859_cast_fp16)[name = string("op_2603_cast_fp16")];
            tensor<int32, [2]> var_2604_split_sizes_0 = const()[name = string("op_2604_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2604_axis_0 = const()[name = string("op_2604_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2604_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2604_cast_fp16_1 = split(axis = var_2604_axis_0, split_sizes = var_2604_split_sizes_0, x = x_51_cast_fp16)[name = string("op_2604_cast_fp16")];
            fp16 const_67_promoted_to_fp16 = const()[name = string("const_67_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_2606_cast_fp16 = mul(x = var_2604_cast_fp16_1, y = const_67_promoted_to_fp16)[name = string("op_2606_cast_fp16")];
            int32 var_2608 = const()[name = string("op_2608"), val = int32(-2)];
            bool var_2609_interleave_0 = const()[name = string("op_2609_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2609_cast_fp16 = concat(axis = var_2608, interleave = var_2609_interleave_0, values = (var_2606_cast_fp16, var_2604_cast_fp16_0))[name = string("op_2609_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2610_cast_fp16 = mul(x = var_2609_cast_fp16, y = var_866_cast_fp16)[name = string("op_2610_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_33_cast_fp16 = add(x = var_2603_cast_fp16, y = var_2610_cast_fp16)[name = string("query_states_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2616_cast_fp16 = mul(x = var_2592_cast_fp16, y = var_859_cast_fp16)[name = string("op_2616_cast_fp16")];
            tensor<int32, [2]> var_2617_split_sizes_0 = const()[name = string("op_2617_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2617_axis_0 = const()[name = string("op_2617_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2617_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2617_cast_fp16_1 = split(axis = var_2617_axis_0, split_sizes = var_2617_split_sizes_0, x = var_2592_cast_fp16)[name = string("op_2617_cast_fp16")];
            fp16 const_68_promoted_to_fp16 = const()[name = string("const_68_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_2619_cast_fp16 = mul(x = var_2617_cast_fp16_1, y = const_68_promoted_to_fp16)[name = string("op_2619_cast_fp16")];
            int32 var_2621 = const()[name = string("op_2621"), val = int32(-2)];
            bool var_2622_interleave_0 = const()[name = string("op_2622_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2622_cast_fp16 = concat(axis = var_2621, interleave = var_2622_interleave_0, values = (var_2619_cast_fp16, var_2617_cast_fp16_0))[name = string("op_2622_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2623_cast_fp16 = mul(x = var_2622_cast_fp16, y = var_866_cast_fp16)[name = string("op_2623_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_55_cast_fp16 = add(x = var_2616_cast_fp16, y = var_2623_cast_fp16)[name = string("key_states_55_cast_fp16")];
            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, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([5])];
            tensor<int32, [1]> expand_dims_62 = const()[name = string("expand_dims_62"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_64 = const()[name = string("expand_dims_64"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_65 = const()[name = string("expand_dims_65"), val = tensor<int32, [1]>([6])];
            int32 concat_43_axis_0 = const()[name = string("concat_43_axis_0"), val = int32(0)];
            bool concat_43_interleave_0 = const()[name = string("concat_43_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_43 = concat(axis = concat_43_axis_0, interleave = concat_43_interleave_0, values = (expand_dims_61, expand_dims_62, position_id, expand_dims_64))[name = string("concat_43")];
            tensor<int32, [1]> concat_44_values1_0 = const()[name = string("concat_44_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_44_values3_0 = const()[name = string("concat_44_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_44_axis_0 = const()[name = string("concat_44_axis_0"), val = int32(0)];
            bool concat_44_interleave_0 = const()[name = string("concat_44_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_44 = concat(axis = concat_44_axis_0, interleave = concat_44_interleave_0, values = (expand_dims_65, concat_44_values1_0, cache_position_end, concat_44_values3_0))[name = string("concat_44")];
            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_67")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_43, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_44, 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, [28, 2, 2048, 128]> coreml_update_state_66 = read_state(input = key_cache)[name = string("coreml_update_state_66")];
            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_2599_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_43, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_44, 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, [28, 2, 2048, 128]> coreml_update_state_67 = read_state(input = value_cache)[name = string("coreml_update_state_67")];
            tensor<int32, [4]> var_2693_begin_0 = const()[name = string("op_2693_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2693_end_0 = const()[name = string("op_2693_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2693_end_mask_0 = const()[name = string("op_2693_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2693_cast_fp16 = slice_by_index(begin = var_2693_begin_0, end = var_2693_end_0, end_mask = var_2693_end_mask_0, x = coreml_update_state_66)[name = string("op_2693_cast_fp16")];
            tensor<int32, [2]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2696_axis_0 = const()[name = string("op_2696_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2696_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2696_cast_fp16_1 = split(axis = var_2696_axis_0, split_sizes = tile_10, x = var_2693_cast_fp16)[name = string("op_2696_cast_fp16")];
            tensor<int32, [4]> var_2703_begin_0 = const()[name = string("op_2703_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_2703_end_0 = const()[name = string("op_2703_end_0"), val = tensor<int32, [4]>([6, 2, 2048, 128])];
            tensor<bool, [4]> var_2703_end_mask_0 = const()[name = string("op_2703_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_2703_cast_fp16 = slice_by_index(begin = var_2703_begin_0, end = var_2703_end_0, end_mask = var_2703_end_mask_0, x = coreml_update_state_67)[name = string("op_2703_cast_fp16")];
            tensor<int32, [2]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2706_axis_0 = const()[name = string("op_2706_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_2706_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_2706_cast_fp16_1 = split(axis = var_2706_axis_0, split_sizes = tile_11, x = var_2703_cast_fp16)[name = string("op_2706_cast_fp16")];
            tensor<int32, [2]> var_2709_split_sizes_0 = const()[name = string("op_2709_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_2709_axis_0 = const()[name = string("op_2709_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_2709_0, tensor<fp16, [1, 8, 128, 8]> var_2709_1 = split(axis = var_2709_axis_0, split_sizes = var_2709_split_sizes_0, x = query_states_33_cast_fp16)[name = string("op_2709")];
            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_2696_cast_fp16_0, y = var_2709_0)[name = string("attn_weights_81_cast_fp16")];
            fp16 var_2712_to_fp16 = const()[name = string("op_2712_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_2712_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_2716 = const()[name = string("op_2716"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_87_cast_fp16 = softmax(axis = var_2716, x = attn_weights_85_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            bool var_2722_transpose_x_1 = const()[name = string("op_2722_transpose_x_1"), val = bool(true)];
            bool var_2722_transpose_y_1 = const()[name = string("op_2722_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_2722_cast_fp16 = matmul(transpose_x = var_2722_transpose_x_1, transpose_y = var_2722_transpose_y_1, x = attn_weights_87_cast_fp16, y = var_2706_cast_fp16_0)[name = string("op_2722_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_2696_cast_fp16_1, y = var_2709_1)[name = string("attn_weights_89_cast_fp16")];
            fp16 var_2724_to_fp16 = const()[name = string("op_2724_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_2724_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_2728 = const()[name = string("op_2728"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_95_cast_fp16 = softmax(axis = var_2728, 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_2706_cast_fp16_1)[name = string("attn_output_41_cast_fp16")];
            int32 var_2736 = const()[name = string("op_2736"), 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_2736, interleave = attn_output_43_interleave_0, values = (var_2722_cast_fp16, attn_output_41_cast_fp16))[name = string("attn_output_43_cast_fp16")];
            tensor<int32, [4]> var_2740_perm_0 = const()[name = string("op_2740_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_2745 = const()[name = string("op_2745"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_2740_cast_fp16 = transpose(perm = var_2740_perm_0, x = attn_output_43_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_47_cast_fp16 = reshape(shape = var_2745, x = var_2740_cast_fp16)[name = string("attn_output_47_cast_fp16")];
            string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_2771 = const()[name = string("op_2771"), val = int32(1)];
            fp16 const_73_promoted_to_fp16 = const()[name = string("const_73_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2773_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = const_73_promoted_to_fp16)[name = string("op_2773_cast_fp16")];
            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_2771, interleave = doubled_45_interleave_0, values = (hidden_states_55_cast_fp16, var_2773_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(1284235520)))];
            fp16 var_2783_to_fp16 = const()[name = string("op_2783_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_2783_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [2]> var_2794_split_sizes_0 = const()[name = string("op_2794_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2794_axis_0 = const()[name = string("op_2794_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2794_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2794_cast_fp16_1 = split(axis = var_2794_axis_0, split_sizes = var_2794_split_sizes_0, x = out_23_cast_fp16)[name = string("op_2794_cast_fp16")];
            string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1284243776)))];
            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_2794_cast_fp16_0)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_2811_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_2811_cast_fp16")];
            string var_2817_pad_type_0 = const()[name = string("op_2817_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2817_strides_0 = const()[name = string("op_2817_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2817_pad_0 = const()[name = string("op_2817_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2817_dilations_0 = const()[name = string("op_2817_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2817_groups_0 = const()[name = string("op_2817_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_2817_cast_fp16 = conv(dilations = var_2817_dilations_0, groups = var_2817_groups_0, pad = var_2817_pad_0, pad_type = var_2817_pad_type_0, strides = var_2817_strides_0, weight = layers_5_mlp_up_proj_weight_cast_fp16, x = var_2794_cast_fp16_0)[name = string("op_2817_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_59_cast_fp16 = mul(x = var_2811_cast_fp16, y = var_2817_cast_fp16)[name = string("x_59_cast_fp16")];
            string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_2833 = const()[name = string("op_2833"), val = int32(1)];
            fp16 const_75_promoted_to_fp16 = const()[name = string("const_75_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_2835_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = const_75_promoted_to_fp16)[name = string("op_2835_cast_fp16")];
            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_2833, interleave = doubled_49_interleave_0, values = (hidden_states_59_cast_fp16, var_2835_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(1309409664)))];
            fp16 var_2845_to_fp16 = const()[name = string("op_2845_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_2845_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [2]> var_2856_split_sizes_0 = const()[name = string("op_2856_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_2856_axis_0 = const()[name = string("op_2856_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_2856_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_2856_cast_fp16_1 = split(axis = var_2856_axis_0, split_sizes = var_2856_split_sizes_0, x = out_25_cast_fp16)[name = string("op_2856_cast_fp16")];
            string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1309417920)))];
            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_2856_cast_fp16_0)[name = string("query_states_37_cast_fp16")];
            string key_states_61_pad_type_0 = const()[name = string("key_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_61_strides_0 = const()[name = string("key_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1317806592)))];
            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_2856_cast_fp16_0)[name = string("key_states_61_cast_fp16")];
            string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_2856_cast_fp16_0)[name = string("value_states_37_cast_fp16")];
            tensor<int32, [4]> var_2908 = const()[name = string("op_2908"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_61_cast_fp16 = reshape(shape = var_2908, x = query_states_37_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> var_2912 = const()[name = string("op_2912"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_2913_cast_fp16 = reshape(shape = var_2912, x = key_states_61_cast_fp16)[name = string("op_2913_cast_fp16")];
            tensor<int32, [4]> var_2919 = const()[name = string("op_2919"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_2920_cast_fp16 = reshape(shape = var_2919, x = value_states_37_cast_fp16)[name = string("op_2920_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_2924_cast_fp16 = mul(x = x_61_cast_fp16, y = var_859_cast_fp16)[name = string("op_2924_cast_fp16")];
            tensor<int32, [2]> var_2925_split_sizes_0 = const()[name = string("op_2925_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2925_axis_0 = const()[name = string("op_2925_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_2925_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_2925_cast_fp16_1 = split(axis = var_2925_axis_0, split_sizes = var_2925_split_sizes_0, x = x_61_cast_fp16)[name = string("op_2925_cast_fp16")];
            fp16 const_79_promoted_to_fp16 = const()[name = string("const_79_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_2927_cast_fp16 = mul(x = var_2925_cast_fp16_1, y = const_79_promoted_to_fp16)[name = string("op_2927_cast_fp16")];
            int32 var_2929 = const()[name = string("op_2929"), val = int32(-2)];
            bool var_2930_interleave_0 = const()[name = string("op_2930_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_2930_cast_fp16 = concat(axis = var_2929, interleave = var_2930_interleave_0, values = (var_2927_cast_fp16, var_2925_cast_fp16_0))[name = string("op_2930_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_2931_cast_fp16 = mul(x = var_2930_cast_fp16, y = var_866_cast_fp16)[name = string("op_2931_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_39_cast_fp16 = add(x = var_2924_cast_fp16, y = var_2931_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2937_cast_fp16 = mul(x = var_2913_cast_fp16, y = var_859_cast_fp16)[name = string("op_2937_cast_fp16")];
            tensor<int32, [2]> var_2938_split_sizes_0 = const()[name = string("op_2938_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2938_axis_0 = const()[name = string("op_2938_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_2938_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_2938_cast_fp16_1 = split(axis = var_2938_axis_0, split_sizes = var_2938_split_sizes_0, x = var_2913_cast_fp16)[name = string("op_2938_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_2940_cast_fp16 = mul(x = var_2938_cast_fp16_1, y = const_80_promoted_to_fp16)[name = string("op_2940_cast_fp16")];
            int32 var_2942 = const()[name = string("op_2942"), val = int32(-2)];
            bool var_2943_interleave_0 = const()[name = string("op_2943_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_2943_cast_fp16 = concat(axis = var_2942, interleave = var_2943_interleave_0, values = (var_2940_cast_fp16, var_2938_cast_fp16_0))[name = string("op_2943_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_2944_cast_fp16 = mul(x = var_2943_cast_fp16, y = var_866_cast_fp16)[name = string("op_2944_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_65_cast_fp16 = add(x = var_2937_cast_fp16, y = var_2944_cast_fp16)[name = string("key_states_65_cast_fp16")];
            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, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([6])];
            tensor<int32, [1]> expand_dims_74 = const()[name = string("expand_dims_74"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_76 = const()[name = string("expand_dims_76"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_77 = const()[name = string("expand_dims_77"), val = tensor<int32, [1]>([7])];
            int32 concat_51_axis_0 = const()[name = string("concat_51_axis_0"), val = int32(0)];
            bool concat_51_interleave_0 = const()[name = string("concat_51_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_51 = concat(axis = concat_51_axis_0, interleave = concat_51_interleave_0, values = (expand_dims_73, expand_dims_74, position_id, expand_dims_76))[name = string("concat_51")];
            tensor<int32, [1]> concat_52_values1_0 = const()[name = string("concat_52_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_52_values3_0 = const()[name = string("concat_52_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_52_axis_0 = const()[name = string("concat_52_axis_0"), val = int32(0)];
            bool concat_52_interleave_0 = const()[name = string("concat_52_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_52 = concat(axis = concat_52_axis_0, interleave = concat_52_interleave_0, values = (expand_dims_77, concat_52_values1_0, cache_position_end, concat_52_values3_0))[name = string("concat_52")];
            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_64")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_51, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_52, 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, [28, 2, 2048, 128]> coreml_update_state_68 = read_state(input = key_cache)[name = string("coreml_update_state_68")];
            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_2920_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_51, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_52, 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, [28, 2, 2048, 128]> coreml_update_state_69 = read_state(input = value_cache)[name = string("coreml_update_state_69")];
            tensor<int32, [4]> var_3014_begin_0 = const()[name = string("op_3014_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3014_end_0 = const()[name = string("op_3014_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_3014_end_mask_0 = const()[name = string("op_3014_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3014_cast_fp16 = slice_by_index(begin = var_3014_begin_0, end = var_3014_end_0, end_mask = var_3014_end_mask_0, x = coreml_update_state_68)[name = string("op_3014_cast_fp16")];
            tensor<int32, [2]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3017_axis_0 = const()[name = string("op_3017_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3017_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3017_cast_fp16_1 = split(axis = var_3017_axis_0, split_sizes = tile_12, x = var_3014_cast_fp16)[name = string("op_3017_cast_fp16")];
            tensor<int32, [4]> var_3024_begin_0 = const()[name = string("op_3024_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3024_end_0 = const()[name = string("op_3024_end_0"), val = tensor<int32, [4]>([7, 2, 2048, 128])];
            tensor<bool, [4]> var_3024_end_mask_0 = const()[name = string("op_3024_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3024_cast_fp16 = slice_by_index(begin = var_3024_begin_0, end = var_3024_end_0, end_mask = var_3024_end_mask_0, x = coreml_update_state_69)[name = string("op_3024_cast_fp16")];
            tensor<int32, [2]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3027_axis_0 = const()[name = string("op_3027_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3027_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3027_cast_fp16_1 = split(axis = var_3027_axis_0, split_sizes = tile_13, x = var_3024_cast_fp16)[name = string("op_3027_cast_fp16")];
            tensor<int32, [2]> var_3030_split_sizes_0 = const()[name = string("op_3030_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3030_axis_0 = const()[name = string("op_3030_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_3030_0, tensor<fp16, [1, 8, 128, 8]> var_3030_1 = split(axis = var_3030_axis_0, split_sizes = var_3030_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_3030")];
            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_3017_cast_fp16_0, y = var_3030_0)[name = string("attn_weights_97_cast_fp16")];
            fp16 var_3033_to_fp16 = const()[name = string("op_3033_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_3033_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_3037 = const()[name = string("op_3037"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_103_cast_fp16 = softmax(axis = var_3037, x = attn_weights_101_cast_fp16)[name = string("attn_weights_103_cast_fp16")];
            bool var_3043_transpose_x_1 = const()[name = string("op_3043_transpose_x_1"), val = bool(true)];
            bool var_3043_transpose_y_1 = const()[name = string("op_3043_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_3043_cast_fp16 = matmul(transpose_x = var_3043_transpose_x_1, transpose_y = var_3043_transpose_y_1, x = attn_weights_103_cast_fp16, y = var_3027_cast_fp16_0)[name = string("op_3043_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_3017_cast_fp16_1, y = var_3030_1)[name = string("attn_weights_105_cast_fp16")];
            fp16 var_3045_to_fp16 = const()[name = string("op_3045_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_3045_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_3049 = const()[name = string("op_3049"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_111_cast_fp16 = softmax(axis = var_3049, 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_3027_cast_fp16_1)[name = string("attn_output_49_cast_fp16")];
            int32 var_3057 = const()[name = string("op_3057"), 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_3057, interleave = attn_output_51_interleave_0, values = (var_3043_cast_fp16, attn_output_49_cast_fp16))[name = string("attn_output_51_cast_fp16")];
            tensor<int32, [4]> var_3061_perm_0 = const()[name = string("op_3061_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_3066 = const()[name = string("op_3066"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_3061_cast_fp16 = transpose(perm = var_3061_perm_0, x = attn_output_51_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_55_cast_fp16 = reshape(shape = var_3066, x = var_3061_cast_fp16)[name = string("attn_output_55_cast_fp16")];
            string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_3092 = const()[name = string("op_3092"), val = int32(1)];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3094_cast_fp16 = mul(x = hidden_states_65_cast_fp16, y = const_85_promoted_to_fp16)[name = string("op_3094_cast_fp16")];
            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_3092, interleave = doubled_53_interleave_0, values = (hidden_states_65_cast_fp16, var_3094_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(1318855232)))];
            fp16 var_3104_to_fp16 = const()[name = string("op_3104_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_3104_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [2]> var_3115_split_sizes_0 = const()[name = string("op_3115_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3115_axis_0 = const()[name = string("op_3115_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3115_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3115_cast_fp16_1 = split(axis = var_3115_axis_0, split_sizes = var_3115_split_sizes_0, x = out_27_cast_fp16)[name = string("op_3115_cast_fp16")];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_3115_cast_fp16_0)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_3132_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_3132_cast_fp16")];
            string var_3138_pad_type_0 = const()[name = string("op_3138_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3138_strides_0 = const()[name = string("op_3138_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3138_pad_0 = const()[name = string("op_3138_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3138_dilations_0 = const()[name = string("op_3138_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3138_groups_0 = const()[name = string("op_3138_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_3138_cast_fp16 = conv(dilations = var_3138_dilations_0, groups = var_3138_groups_0, pad = var_3138_pad_0, pad_type = var_3138_pad_type_0, strides = var_3138_strides_0, weight = layers_6_mlp_up_proj_weight_cast_fp16, x = var_3115_cast_fp16_0)[name = string("op_3138_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_69_cast_fp16 = mul(x = var_3132_cast_fp16, y = var_3138_cast_fp16)[name = string("x_69_cast_fp16")];
            string hidden_states_67_pad_type_0 = const()[name = string("hidden_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_67_strides_0 = const()[name = string("hidden_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_3154 = const()[name = string("op_3154"), val = int32(1)];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3156_cast_fp16 = mul(x = hidden_states_69_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_3156_cast_fp16")];
            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_3154, interleave = doubled_57_interleave_0, values = (hidden_states_69_cast_fp16, var_3156_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(1318863488)))];
            fp16 var_3166_to_fp16 = const()[name = string("op_3166_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_3166_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [2]> var_3177_split_sizes_0 = const()[name = string("op_3177_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3177_axis_0 = const()[name = string("op_3177_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3177_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3177_cast_fp16_1 = split(axis = var_3177_axis_0, split_sizes = var_3177_split_sizes_0, x = out_29_cast_fp16)[name = string("op_3177_cast_fp16")];
            string query_states_43_pad_type_0 = const()[name = string("query_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_43_strides_0 = const()[name = string("query_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1318871744)))];
            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_3177_cast_fp16_0)[name = string("query_states_43_cast_fp16")];
            string key_states_71_pad_type_0 = const()[name = string("key_states_71_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_71_strides_0 = const()[name = string("key_states_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1327260416)))];
            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_3177_cast_fp16_0)[name = string("key_states_71_cast_fp16")];
            string value_states_43_pad_type_0 = const()[name = string("value_states_43_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_43_strides_0 = const()[name = string("value_states_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_3177_cast_fp16_0)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> var_3229 = const()[name = string("op_3229"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_71_cast_fp16 = reshape(shape = var_3229, x = query_states_43_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> var_3233 = const()[name = string("op_3233"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_3234_cast_fp16 = reshape(shape = var_3233, x = key_states_71_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [4]> var_3240 = const()[name = string("op_3240"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_3241_cast_fp16 = reshape(shape = var_3240, x = value_states_43_cast_fp16)[name = string("op_3241_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_3245_cast_fp16 = mul(x = x_71_cast_fp16, y = var_859_cast_fp16)[name = string("op_3245_cast_fp16")];
            tensor<int32, [2]> var_3246_split_sizes_0 = const()[name = string("op_3246_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3246_axis_0 = const()[name = string("op_3246_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_3246_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_3246_cast_fp16_1 = split(axis = var_3246_axis_0, split_sizes = var_3246_split_sizes_0, x = x_71_cast_fp16)[name = string("op_3246_cast_fp16")];
            fp16 const_91_promoted_to_fp16 = const()[name = string("const_91_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_3248_cast_fp16 = mul(x = var_3246_cast_fp16_1, y = const_91_promoted_to_fp16)[name = string("op_3248_cast_fp16")];
            int32 var_3250 = const()[name = string("op_3250"), val = int32(-2)];
            bool var_3251_interleave_0 = const()[name = string("op_3251_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_3251_cast_fp16 = concat(axis = var_3250, interleave = var_3251_interleave_0, values = (var_3248_cast_fp16, var_3246_cast_fp16_0))[name = string("op_3251_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3252_cast_fp16 = mul(x = var_3251_cast_fp16, y = var_866_cast_fp16)[name = string("op_3252_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_45_cast_fp16 = add(x = var_3245_cast_fp16, y = var_3252_cast_fp16)[name = string("query_states_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3258_cast_fp16 = mul(x = var_3234_cast_fp16, y = var_859_cast_fp16)[name = string("op_3258_cast_fp16")];
            tensor<int32, [2]> var_3259_split_sizes_0 = const()[name = string("op_3259_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3259_axis_0 = const()[name = string("op_3259_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_3259_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_3259_cast_fp16_1 = split(axis = var_3259_axis_0, split_sizes = var_3259_split_sizes_0, x = var_3234_cast_fp16)[name = string("op_3259_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_3261_cast_fp16 = mul(x = var_3259_cast_fp16_1, y = const_92_promoted_to_fp16)[name = string("op_3261_cast_fp16")];
            int32 var_3263 = const()[name = string("op_3263"), val = int32(-2)];
            bool var_3264_interleave_0 = const()[name = string("op_3264_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_3264_cast_fp16 = concat(axis = var_3263, interleave = var_3264_interleave_0, values = (var_3261_cast_fp16, var_3259_cast_fp16_0))[name = string("op_3264_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3265_cast_fp16 = mul(x = var_3264_cast_fp16, y = var_866_cast_fp16)[name = string("op_3265_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_75_cast_fp16 = add(x = var_3258_cast_fp16, y = var_3265_cast_fp16)[name = string("key_states_75_cast_fp16")];
            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, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([7])];
            tensor<int32, [1]> expand_dims_86 = const()[name = string("expand_dims_86"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_88 = const()[name = string("expand_dims_88"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_89 = const()[name = string("expand_dims_89"), val = tensor<int32, [1]>([8])];
            int32 concat_59_axis_0 = const()[name = string("concat_59_axis_0"), val = int32(0)];
            bool concat_59_interleave_0 = const()[name = string("concat_59_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_59 = concat(axis = concat_59_axis_0, interleave = concat_59_interleave_0, values = (expand_dims_85, expand_dims_86, position_id, expand_dims_88))[name = string("concat_59")];
            tensor<int32, [1]> concat_60_values1_0 = const()[name = string("concat_60_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_60_values3_0 = const()[name = string("concat_60_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_60_axis_0 = const()[name = string("concat_60_axis_0"), val = int32(0)];
            bool concat_60_interleave_0 = const()[name = string("concat_60_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_60 = concat(axis = concat_60_axis_0, interleave = concat_60_interleave_0, values = (expand_dims_89, concat_60_values1_0, cache_position_end, concat_60_values3_0))[name = string("concat_60")];
            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_61")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_59, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_60, 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, [28, 2, 2048, 128]> coreml_update_state_70 = read_state(input = key_cache)[name = string("coreml_update_state_70")];
            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_3241_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_59, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_60, 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, [28, 2, 2048, 128]> coreml_update_state_71 = read_state(input = value_cache)[name = string("coreml_update_state_71")];
            tensor<int32, [4]> var_3335_begin_0 = const()[name = string("op_3335_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_3335_end_0 = const()[name = string("op_3335_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_3335_end_mask_0 = const()[name = string("op_3335_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3335_cast_fp16 = slice_by_index(begin = var_3335_begin_0, end = var_3335_end_0, end_mask = var_3335_end_mask_0, x = coreml_update_state_70)[name = string("op_3335_cast_fp16")];
            tensor<int32, [2]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3338_axis_0 = const()[name = string("op_3338_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3338_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3338_cast_fp16_1 = split(axis = var_3338_axis_0, split_sizes = tile_14, x = var_3335_cast_fp16)[name = string("op_3338_cast_fp16")];
            tensor<int32, [4]> var_3345_begin_0 = const()[name = string("op_3345_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_3345_end_0 = const()[name = string("op_3345_end_0"), val = tensor<int32, [4]>([8, 2, 2048, 128])];
            tensor<bool, [4]> var_3345_end_mask_0 = const()[name = string("op_3345_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3345_cast_fp16 = slice_by_index(begin = var_3345_begin_0, end = var_3345_end_0, end_mask = var_3345_end_mask_0, x = coreml_update_state_71)[name = string("op_3345_cast_fp16")];
            tensor<int32, [2]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3348_axis_0 = const()[name = string("op_3348_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3348_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3348_cast_fp16_1 = split(axis = var_3348_axis_0, split_sizes = tile_15, x = var_3345_cast_fp16)[name = string("op_3348_cast_fp16")];
            tensor<int32, [2]> var_3351_split_sizes_0 = const()[name = string("op_3351_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3351_axis_0 = const()[name = string("op_3351_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_3351_0, tensor<fp16, [1, 8, 128, 8]> var_3351_1 = split(axis = var_3351_axis_0, split_sizes = var_3351_split_sizes_0, x = query_states_45_cast_fp16)[name = string("op_3351")];
            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_3338_cast_fp16_0, y = var_3351_0)[name = string("attn_weights_113_cast_fp16")];
            fp16 var_3354_to_fp16 = const()[name = string("op_3354_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_3354_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_3358 = const()[name = string("op_3358"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_119_cast_fp16 = softmax(axis = var_3358, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")];
            bool var_3364_transpose_x_1 = const()[name = string("op_3364_transpose_x_1"), val = bool(true)];
            bool var_3364_transpose_y_1 = const()[name = string("op_3364_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_3364_cast_fp16 = matmul(transpose_x = var_3364_transpose_x_1, transpose_y = var_3364_transpose_y_1, x = attn_weights_119_cast_fp16, y = var_3348_cast_fp16_0)[name = string("op_3364_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_3338_cast_fp16_1, y = var_3351_1)[name = string("attn_weights_121_cast_fp16")];
            fp16 var_3366_to_fp16 = const()[name = string("op_3366_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_3366_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_3370 = const()[name = string("op_3370"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_127_cast_fp16 = softmax(axis = var_3370, 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_3348_cast_fp16_1)[name = string("attn_output_57_cast_fp16")];
            int32 var_3378 = const()[name = string("op_3378"), 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_3378, interleave = attn_output_59_interleave_0, values = (var_3364_cast_fp16, attn_output_57_cast_fp16))[name = string("attn_output_59_cast_fp16")];
            tensor<int32, [4]> var_3382_perm_0 = const()[name = string("op_3382_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_3387 = const()[name = string("op_3387"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_3382_cast_fp16 = transpose(perm = var_3382_perm_0, x = attn_output_59_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_63_cast_fp16 = reshape(shape = var_3387, x = var_3382_cast_fp16)[name = string("attn_output_63_cast_fp16")];
            string hidden_states_73_pad_type_0 = const()[name = string("hidden_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_73_strides_0 = const()[name = string("hidden_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_3413 = const()[name = string("op_3413"), val = int32(1)];
            fp16 const_97_promoted_to_fp16 = const()[name = string("const_97_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3415_cast_fp16 = mul(x = hidden_states_75_cast_fp16, y = const_97_promoted_to_fp16)[name = string("op_3415_cast_fp16")];
            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_3413, interleave = doubled_61_interleave_0, values = (hidden_states_75_cast_fp16, var_3415_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(1328309056)))];
            fp16 var_3425_to_fp16 = const()[name = string("op_3425_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_3425_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [2]> var_3436_split_sizes_0 = const()[name = string("op_3436_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3436_axis_0 = const()[name = string("op_3436_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3436_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3436_cast_fp16_1 = split(axis = var_3436_axis_0, split_sizes = var_3436_split_sizes_0, x = out_31_cast_fp16)[name = string("op_3436_cast_fp16")];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_3436_cast_fp16_0)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_3453_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_3453_cast_fp16")];
            string var_3459_pad_type_0 = const()[name = string("op_3459_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3459_strides_0 = const()[name = string("op_3459_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3459_pad_0 = const()[name = string("op_3459_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3459_dilations_0 = const()[name = string("op_3459_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3459_groups_0 = const()[name = string("op_3459_groups_0"), val = int32(1)];
            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(1328317312)))];
            tensor<fp16, [1, 6144, 1, 8]> var_3459_cast_fp16 = conv(dilations = var_3459_dilations_0, groups = var_3459_groups_0, pad = var_3459_pad_0, pad_type = var_3459_pad_type_0, strides = var_3459_strides_0, weight = layers_7_mlp_up_proj_weight_to_fp16, x = var_3436_cast_fp16_0)[name = string("op_3459_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_79_cast_fp16 = mul(x = var_3453_cast_fp16, y = var_3459_cast_fp16)[name = string("x_79_cast_fp16")];
            string hidden_states_77_pad_type_0 = const()[name = string("hidden_states_77_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_77_strides_0 = const()[name = string("hidden_states_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1353483200)))];
            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")];
            int32 var_3475 = const()[name = string("op_3475"), val = int32(1)];
            fp16 const_99_promoted_to_fp16 = const()[name = string("const_99_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3477_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = const_99_promoted_to_fp16)[name = string("op_3477_cast_fp16")];
            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_3475, interleave = doubled_65_interleave_0, values = (hidden_states_79_cast_fp16, var_3477_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(1378649088)))];
            fp16 var_3487_to_fp16 = const()[name = string("op_3487_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_3487_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [2]> var_3498_split_sizes_0 = const()[name = string("op_3498_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3498_axis_0 = const()[name = string("op_3498_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3498_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3498_cast_fp16_1 = split(axis = var_3498_axis_0, split_sizes = var_3498_split_sizes_0, x = out_33_cast_fp16)[name = string("op_3498_cast_fp16")];
            string query_states_49_pad_type_0 = const()[name = string("query_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_49_strides_0 = const()[name = string("query_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1378657344)))];
            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_3498_cast_fp16_0)[name = string("query_states_49_cast_fp16")];
            string key_states_81_pad_type_0 = const()[name = string("key_states_81_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_81_strides_0 = const()[name = string("key_states_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1387046016)))];
            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_3498_cast_fp16_0)[name = string("key_states_81_cast_fp16")];
            string value_states_49_pad_type_0 = const()[name = string("value_states_49_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_49_strides_0 = const()[name = string("value_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_3498_cast_fp16_0)[name = string("value_states_49_cast_fp16")];
            tensor<int32, [4]> var_3550 = const()[name = string("op_3550"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_81_cast_fp16 = reshape(shape = var_3550, x = query_states_49_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> var_3554 = const()[name = string("op_3554"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_3555_cast_fp16 = reshape(shape = var_3554, x = key_states_81_cast_fp16)[name = string("op_3555_cast_fp16")];
            tensor<int32, [4]> var_3561 = const()[name = string("op_3561"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_3562_cast_fp16 = reshape(shape = var_3561, x = value_states_49_cast_fp16)[name = string("op_3562_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_3566_cast_fp16 = mul(x = x_81_cast_fp16, y = var_859_cast_fp16)[name = string("op_3566_cast_fp16")];
            tensor<int32, [2]> var_3567_split_sizes_0 = const()[name = string("op_3567_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3567_axis_0 = const()[name = string("op_3567_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_3567_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_3567_cast_fp16_1 = split(axis = var_3567_axis_0, split_sizes = var_3567_split_sizes_0, x = x_81_cast_fp16)[name = string("op_3567_cast_fp16")];
            fp16 const_103_promoted_to_fp16 = const()[name = string("const_103_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_3569_cast_fp16 = mul(x = var_3567_cast_fp16_1, y = const_103_promoted_to_fp16)[name = string("op_3569_cast_fp16")];
            int32 var_3571 = const()[name = string("op_3571"), val = int32(-2)];
            bool var_3572_interleave_0 = const()[name = string("op_3572_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_3572_cast_fp16 = concat(axis = var_3571, interleave = var_3572_interleave_0, values = (var_3569_cast_fp16, var_3567_cast_fp16_0))[name = string("op_3572_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3573_cast_fp16 = mul(x = var_3572_cast_fp16, y = var_866_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_51_cast_fp16 = add(x = var_3566_cast_fp16, y = var_3573_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3579_cast_fp16 = mul(x = var_3555_cast_fp16, y = var_859_cast_fp16)[name = string("op_3579_cast_fp16")];
            tensor<int32, [2]> var_3580_split_sizes_0 = const()[name = string("op_3580_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3580_axis_0 = const()[name = string("op_3580_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_3580_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_3580_cast_fp16_1 = split(axis = var_3580_axis_0, split_sizes = var_3580_split_sizes_0, x = var_3555_cast_fp16)[name = string("op_3580_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_3582_cast_fp16 = mul(x = var_3580_cast_fp16_1, y = const_104_promoted_to_fp16)[name = string("op_3582_cast_fp16")];
            int32 var_3584 = const()[name = string("op_3584"), val = int32(-2)];
            bool var_3585_interleave_0 = const()[name = string("op_3585_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_3585_cast_fp16 = concat(axis = var_3584, interleave = var_3585_interleave_0, values = (var_3582_cast_fp16, var_3580_cast_fp16_0))[name = string("op_3585_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3586_cast_fp16 = mul(x = var_3585_cast_fp16, y = var_866_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_85_cast_fp16 = add(x = var_3579_cast_fp16, y = var_3586_cast_fp16)[name = string("key_states_85_cast_fp16")];
            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, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_98 = const()[name = string("expand_dims_98"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_100 = const()[name = string("expand_dims_100"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_101 = const()[name = string("expand_dims_101"), val = tensor<int32, [1]>([9])];
            int32 concat_67_axis_0 = const()[name = string("concat_67_axis_0"), val = int32(0)];
            bool concat_67_interleave_0 = const()[name = string("concat_67_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_67 = concat(axis = concat_67_axis_0, interleave = concat_67_interleave_0, values = (expand_dims_97, expand_dims_98, position_id, expand_dims_100))[name = string("concat_67")];
            tensor<int32, [1]> concat_68_values1_0 = const()[name = string("concat_68_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_68_values3_0 = const()[name = string("concat_68_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_68_axis_0 = const()[name = string("concat_68_axis_0"), val = int32(0)];
            bool concat_68_interleave_0 = const()[name = string("concat_68_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_68 = concat(axis = concat_68_axis_0, interleave = concat_68_interleave_0, values = (expand_dims_101, concat_68_values1_0, cache_position_end, concat_68_values3_0))[name = string("concat_68")];
            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_58")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_67, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_68, 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, [28, 2, 2048, 128]> coreml_update_state_72 = read_state(input = key_cache)[name = string("coreml_update_state_72")];
            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_3562_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_67, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_68, 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, [28, 2, 2048, 128]> coreml_update_state_73 = read_state(input = value_cache)[name = string("coreml_update_state_73")];
            tensor<int32, [4]> var_3656_begin_0 = const()[name = string("op_3656_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3656_end_0 = const()[name = string("op_3656_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3656_end_mask_0 = const()[name = string("op_3656_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3656_cast_fp16 = slice_by_index(begin = var_3656_begin_0, end = var_3656_end_0, end_mask = var_3656_end_mask_0, x = coreml_update_state_72)[name = string("op_3656_cast_fp16")];
            tensor<int32, [2]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3659_axis_0 = const()[name = string("op_3659_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3659_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3659_cast_fp16_1 = split(axis = var_3659_axis_0, split_sizes = tile_16, x = var_3656_cast_fp16)[name = string("op_3659_cast_fp16")];
            tensor<int32, [4]> var_3666_begin_0 = const()[name = string("op_3666_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_3666_end_0 = const()[name = string("op_3666_end_0"), val = tensor<int32, [4]>([9, 2, 2048, 128])];
            tensor<bool, [4]> var_3666_end_mask_0 = const()[name = string("op_3666_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3666_cast_fp16 = slice_by_index(begin = var_3666_begin_0, end = var_3666_end_0, end_mask = var_3666_end_mask_0, x = coreml_update_state_73)[name = string("op_3666_cast_fp16")];
            tensor<int32, [2]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3669_axis_0 = const()[name = string("op_3669_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3669_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3669_cast_fp16_1 = split(axis = var_3669_axis_0, split_sizes = tile_17, x = var_3666_cast_fp16)[name = string("op_3669_cast_fp16")];
            tensor<int32, [2]> var_3672_split_sizes_0 = const()[name = string("op_3672_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3672_axis_0 = const()[name = string("op_3672_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_3672_0, tensor<fp16, [1, 8, 128, 8]> var_3672_1 = split(axis = var_3672_axis_0, split_sizes = var_3672_split_sizes_0, x = query_states_51_cast_fp16)[name = string("op_3672")];
            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_3659_cast_fp16_0, y = var_3672_0)[name = string("attn_weights_129_cast_fp16")];
            fp16 var_3675_to_fp16 = const()[name = string("op_3675_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_3675_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_3679 = const()[name = string("op_3679"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_135_cast_fp16 = softmax(axis = var_3679, x = attn_weights_133_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            bool var_3685_transpose_x_1 = const()[name = string("op_3685_transpose_x_1"), val = bool(true)];
            bool var_3685_transpose_y_1 = const()[name = string("op_3685_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_3685_cast_fp16 = matmul(transpose_x = var_3685_transpose_x_1, transpose_y = var_3685_transpose_y_1, x = attn_weights_135_cast_fp16, y = var_3669_cast_fp16_0)[name = string("op_3685_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_3659_cast_fp16_1, y = var_3672_1)[name = string("attn_weights_137_cast_fp16")];
            fp16 var_3687_to_fp16 = const()[name = string("op_3687_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_3687_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_3691 = const()[name = string("op_3691"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_143_cast_fp16 = softmax(axis = var_3691, 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_3669_cast_fp16_1)[name = string("attn_output_65_cast_fp16")];
            int32 var_3699 = const()[name = string("op_3699"), 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_3699, interleave = attn_output_67_interleave_0, values = (var_3685_cast_fp16, attn_output_65_cast_fp16))[name = string("attn_output_67_cast_fp16")];
            tensor<int32, [4]> var_3703_perm_0 = const()[name = string("op_3703_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_3708 = const()[name = string("op_3708"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_3703_cast_fp16 = transpose(perm = var_3703_perm_0, x = attn_output_67_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_71_cast_fp16 = reshape(shape = var_3708, x = var_3703_cast_fp16)[name = string("attn_output_71_cast_fp16")];
            string hidden_states_83_pad_type_0 = const()[name = string("hidden_states_83_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_83_strides_0 = const()[name = string("hidden_states_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_3734 = const()[name = string("op_3734"), val = int32(1)];
            fp16 const_109_promoted_to_fp16 = const()[name = string("const_109_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3736_cast_fp16 = mul(x = hidden_states_85_cast_fp16, y = const_109_promoted_to_fp16)[name = string("op_3736_cast_fp16")];
            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_3734, interleave = doubled_69_interleave_0, values = (hidden_states_85_cast_fp16, var_3736_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(1388094656)))];
            fp16 var_3746_to_fp16 = const()[name = string("op_3746_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_3746_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [2]> var_3757_split_sizes_0 = const()[name = string("op_3757_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3757_axis_0 = const()[name = string("op_3757_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3757_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3757_cast_fp16_1 = split(axis = var_3757_axis_0, split_sizes = var_3757_split_sizes_0, x = out_35_cast_fp16)[name = string("op_3757_cast_fp16")];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_3757_cast_fp16_0)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_3774_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_3774_cast_fp16")];
            string var_3780_pad_type_0 = const()[name = string("op_3780_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3780_strides_0 = const()[name = string("op_3780_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3780_pad_0 = const()[name = string("op_3780_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3780_dilations_0 = const()[name = string("op_3780_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3780_groups_0 = const()[name = string("op_3780_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_3780_cast_fp16 = conv(dilations = var_3780_dilations_0, groups = var_3780_groups_0, pad = var_3780_pad_0, pad_type = var_3780_pad_type_0, strides = var_3780_strides_0, weight = layers_8_mlp_up_proj_weight_cast_fp16, x = var_3757_cast_fp16_0)[name = string("op_3780_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_89_cast_fp16 = mul(x = var_3774_cast_fp16, y = var_3780_cast_fp16)[name = string("x_89_cast_fp16")];
            string hidden_states_87_pad_type_0 = const()[name = string("hidden_states_87_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_87_strides_0 = const()[name = string("hidden_states_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_3796 = const()[name = string("op_3796"), val = int32(1)];
            fp16 const_111_promoted_to_fp16 = const()[name = string("const_111_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_3798_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = const_111_promoted_to_fp16)[name = string("op_3798_cast_fp16")];
            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_3796, interleave = doubled_73_interleave_0, values = (hidden_states_89_cast_fp16, var_3798_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(1388102912)))];
            fp16 var_3808_to_fp16 = const()[name = string("op_3808_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_3808_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [2]> var_3819_split_sizes_0 = const()[name = string("op_3819_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_3819_axis_0 = const()[name = string("op_3819_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_3819_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_3819_cast_fp16_1 = split(axis = var_3819_axis_0, split_sizes = var_3819_split_sizes_0, x = out_37_cast_fp16)[name = string("op_3819_cast_fp16")];
            string query_states_55_pad_type_0 = const()[name = string("query_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_55_strides_0 = const()[name = string("query_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1388111168)))];
            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_3819_cast_fp16_0)[name = string("query_states_55_cast_fp16")];
            string key_states_91_pad_type_0 = const()[name = string("key_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_91_strides_0 = const()[name = string("key_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1396499840)))];
            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_3819_cast_fp16_0)[name = string("key_states_91_cast_fp16")];
            string value_states_55_pad_type_0 = const()[name = string("value_states_55_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_55_strides_0 = const()[name = string("value_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_3819_cast_fp16_0)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> var_3871 = const()[name = string("op_3871"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_91_cast_fp16 = reshape(shape = var_3871, x = query_states_55_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> var_3875 = const()[name = string("op_3875"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_3876_cast_fp16 = reshape(shape = var_3875, x = key_states_91_cast_fp16)[name = string("op_3876_cast_fp16")];
            tensor<int32, [4]> var_3882 = const()[name = string("op_3882"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_3883_cast_fp16 = reshape(shape = var_3882, x = value_states_55_cast_fp16)[name = string("op_3883_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_3887_cast_fp16 = mul(x = x_91_cast_fp16, y = var_859_cast_fp16)[name = string("op_3887_cast_fp16")];
            tensor<int32, [2]> var_3888_split_sizes_0 = const()[name = string("op_3888_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3888_axis_0 = const()[name = string("op_3888_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_3888_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_3888_cast_fp16_1 = split(axis = var_3888_axis_0, split_sizes = var_3888_split_sizes_0, x = x_91_cast_fp16)[name = string("op_3888_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_3890_cast_fp16 = mul(x = var_3888_cast_fp16_1, y = const_115_promoted_to_fp16)[name = string("op_3890_cast_fp16")];
            int32 var_3892 = const()[name = string("op_3892"), val = int32(-2)];
            bool var_3893_interleave_0 = const()[name = string("op_3893_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_3893_cast_fp16 = concat(axis = var_3892, interleave = var_3893_interleave_0, values = (var_3890_cast_fp16, var_3888_cast_fp16_0))[name = string("op_3893_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_3894_cast_fp16 = mul(x = var_3893_cast_fp16, y = var_866_cast_fp16)[name = string("op_3894_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_57_cast_fp16 = add(x = var_3887_cast_fp16, y = var_3894_cast_fp16)[name = string("query_states_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3900_cast_fp16 = mul(x = var_3876_cast_fp16, y = var_859_cast_fp16)[name = string("op_3900_cast_fp16")];
            tensor<int32, [2]> var_3901_split_sizes_0 = const()[name = string("op_3901_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3901_axis_0 = const()[name = string("op_3901_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_3901_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_3901_cast_fp16_1 = split(axis = var_3901_axis_0, split_sizes = var_3901_split_sizes_0, x = var_3876_cast_fp16)[name = string("op_3901_cast_fp16")];
            fp16 const_116_promoted_to_fp16 = const()[name = string("const_116_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_3903_cast_fp16 = mul(x = var_3901_cast_fp16_1, y = const_116_promoted_to_fp16)[name = string("op_3903_cast_fp16")];
            int32 var_3905 = const()[name = string("op_3905"), val = int32(-2)];
            bool var_3906_interleave_0 = const()[name = string("op_3906_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_3906_cast_fp16 = concat(axis = var_3905, interleave = var_3906_interleave_0, values = (var_3903_cast_fp16, var_3901_cast_fp16_0))[name = string("op_3906_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_3907_cast_fp16 = mul(x = var_3906_cast_fp16, y = var_866_cast_fp16)[name = string("op_3907_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_95_cast_fp16 = add(x = var_3900_cast_fp16, y = var_3907_cast_fp16)[name = string("key_states_95_cast_fp16")];
            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, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([9])];
            tensor<int32, [1]> expand_dims_110 = const()[name = string("expand_dims_110"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_112 = const()[name = string("expand_dims_112"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_113 = const()[name = string("expand_dims_113"), val = tensor<int32, [1]>([10])];
            int32 concat_75_axis_0 = const()[name = string("concat_75_axis_0"), val = int32(0)];
            bool concat_75_interleave_0 = const()[name = string("concat_75_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_75 = concat(axis = concat_75_axis_0, interleave = concat_75_interleave_0, values = (expand_dims_109, expand_dims_110, position_id, expand_dims_112))[name = string("concat_75")];
            tensor<int32, [1]> concat_76_values1_0 = const()[name = string("concat_76_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_76_values3_0 = const()[name = string("concat_76_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_76_axis_0 = const()[name = string("concat_76_axis_0"), val = int32(0)];
            bool concat_76_interleave_0 = const()[name = string("concat_76_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_76 = concat(axis = concat_76_axis_0, interleave = concat_76_interleave_0, values = (expand_dims_113, concat_76_values1_0, cache_position_end, concat_76_values3_0))[name = string("concat_76")];
            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_55")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_75, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_76, 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, [28, 2, 2048, 128]> coreml_update_state_74 = read_state(input = key_cache)[name = string("coreml_update_state_74")];
            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_3883_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_75, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_76, 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, [28, 2, 2048, 128]> coreml_update_state_75 = read_state(input = value_cache)[name = string("coreml_update_state_75")];
            tensor<int32, [4]> var_3977_begin_0 = const()[name = string("op_3977_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3977_end_0 = const()[name = string("op_3977_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3977_end_mask_0 = const()[name = string("op_3977_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3977_cast_fp16 = slice_by_index(begin = var_3977_begin_0, end = var_3977_end_0, end_mask = var_3977_end_mask_0, x = coreml_update_state_74)[name = string("op_3977_cast_fp16")];
            tensor<int32, [2]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3980_axis_0 = const()[name = string("op_3980_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3980_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3980_cast_fp16_1 = split(axis = var_3980_axis_0, split_sizes = tile_18, x = var_3977_cast_fp16)[name = string("op_3980_cast_fp16")];
            tensor<int32, [4]> var_3987_begin_0 = const()[name = string("op_3987_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_3987_end_0 = const()[name = string("op_3987_end_0"), val = tensor<int32, [4]>([10, 2, 2048, 128])];
            tensor<bool, [4]> var_3987_end_mask_0 = const()[name = string("op_3987_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_3987_cast_fp16 = slice_by_index(begin = var_3987_begin_0, end = var_3987_end_0, end_mask = var_3987_end_mask_0, x = coreml_update_state_75)[name = string("op_3987_cast_fp16")];
            tensor<int32, [2]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3990_axis_0 = const()[name = string("op_3990_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_3990_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_3990_cast_fp16_1 = split(axis = var_3990_axis_0, split_sizes = tile_19, x = var_3987_cast_fp16)[name = string("op_3990_cast_fp16")];
            tensor<int32, [2]> var_3993_split_sizes_0 = const()[name = string("op_3993_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_3993_axis_0 = const()[name = string("op_3993_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_3993_0, tensor<fp16, [1, 8, 128, 8]> var_3993_1 = split(axis = var_3993_axis_0, split_sizes = var_3993_split_sizes_0, x = query_states_57_cast_fp16)[name = string("op_3993")];
            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_3980_cast_fp16_0, y = var_3993_0)[name = string("attn_weights_145_cast_fp16")];
            fp16 var_3996_to_fp16 = const()[name = string("op_3996_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_3996_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_4000 = const()[name = string("op_4000"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_151_cast_fp16 = softmax(axis = var_4000, x = attn_weights_149_cast_fp16)[name = string("attn_weights_151_cast_fp16")];
            bool var_4006_transpose_x_1 = const()[name = string("op_4006_transpose_x_1"), val = bool(true)];
            bool var_4006_transpose_y_1 = const()[name = string("op_4006_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_4006_cast_fp16 = matmul(transpose_x = var_4006_transpose_x_1, transpose_y = var_4006_transpose_y_1, x = attn_weights_151_cast_fp16, y = var_3990_cast_fp16_0)[name = string("op_4006_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_3980_cast_fp16_1, y = var_3993_1)[name = string("attn_weights_153_cast_fp16")];
            fp16 var_4008_to_fp16 = const()[name = string("op_4008_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_4008_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_4012 = const()[name = string("op_4012"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_159_cast_fp16 = softmax(axis = var_4012, 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_3990_cast_fp16_1)[name = string("attn_output_73_cast_fp16")];
            int32 var_4020 = const()[name = string("op_4020"), 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_4020, interleave = attn_output_75_interleave_0, values = (var_4006_cast_fp16, attn_output_73_cast_fp16))[name = string("attn_output_75_cast_fp16")];
            tensor<int32, [4]> var_4024_perm_0 = const()[name = string("op_4024_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_4029 = const()[name = string("op_4029"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_4024_cast_fp16 = transpose(perm = var_4024_perm_0, x = attn_output_75_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_79_cast_fp16 = reshape(shape = var_4029, x = var_4024_cast_fp16)[name = string("attn_output_79_cast_fp16")];
            string hidden_states_93_pad_type_0 = const()[name = string("hidden_states_93_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_93_strides_0 = const()[name = string("hidden_states_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_4055 = const()[name = string("op_4055"), val = int32(1)];
            fp16 const_121_promoted_to_fp16 = const()[name = string("const_121_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4057_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = const_121_promoted_to_fp16)[name = string("op_4057_cast_fp16")];
            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_4055, interleave = doubled_77_interleave_0, values = (hidden_states_95_cast_fp16, var_4057_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(1397548480)))];
            fp16 var_4067_to_fp16 = const()[name = string("op_4067_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_4067_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [2]> var_4078_split_sizes_0 = const()[name = string("op_4078_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4078_axis_0 = const()[name = string("op_4078_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4078_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4078_cast_fp16_1 = split(axis = var_4078_axis_0, split_sizes = var_4078_split_sizes_0, x = out_39_cast_fp16)[name = string("op_4078_cast_fp16")];
            string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_4078_cast_fp16_0)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_4095_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_4095_cast_fp16")];
            string var_4101_pad_type_0 = const()[name = string("op_4101_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4101_strides_0 = const()[name = string("op_4101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4101_pad_0 = const()[name = string("op_4101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4101_dilations_0 = const()[name = string("op_4101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4101_groups_0 = const()[name = string("op_4101_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_4101_cast_fp16 = conv(dilations = var_4101_dilations_0, groups = var_4101_groups_0, pad = var_4101_pad_0, pad_type = var_4101_pad_type_0, strides = var_4101_strides_0, weight = layers_9_mlp_up_proj_weight_cast_fp16, x = var_4078_cast_fp16_0)[name = string("op_4101_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_99_cast_fp16 = mul(x = var_4095_cast_fp16, y = var_4101_cast_fp16)[name = string("x_99_cast_fp16")];
            string hidden_states_97_pad_type_0 = const()[name = string("hidden_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_97_strides_0 = const()[name = string("hidden_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_4117 = const()[name = string("op_4117"), val = int32(1)];
            fp16 const_123_promoted_to_fp16 = const()[name = string("const_123_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4119_cast_fp16 = mul(x = hidden_states_99_cast_fp16, y = const_123_promoted_to_fp16)[name = string("op_4119_cast_fp16")];
            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_4117, interleave = doubled_81_interleave_0, values = (hidden_states_99_cast_fp16, var_4119_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(1397556736)))];
            fp16 var_4129_to_fp16 = const()[name = string("op_4129_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_4129_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [2]> var_4140_split_sizes_0 = const()[name = string("op_4140_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4140_axis_0 = const()[name = string("op_4140_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4140_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4140_cast_fp16_1 = split(axis = var_4140_axis_0, split_sizes = var_4140_split_sizes_0, x = out_41_cast_fp16)[name = string("op_4140_cast_fp16")];
            string query_states_61_pad_type_0 = const()[name = string("query_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_61_strides_0 = const()[name = string("query_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1397564992)))];
            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_4140_cast_fp16_0)[name = string("query_states_61_cast_fp16")];
            string key_states_101_pad_type_0 = const()[name = string("key_states_101_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_101_strides_0 = const()[name = string("key_states_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [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(1405953664)))];
            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_4140_cast_fp16_0)[name = string("key_states_101_cast_fp16")];
            string value_states_61_pad_type_0 = const()[name = string("value_states_61_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_61_strides_0 = const()[name = string("value_states_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_4140_cast_fp16_0)[name = string("value_states_61_cast_fp16")];
            tensor<int32, [4]> var_4192 = const()[name = string("op_4192"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_101_cast_fp16 = reshape(shape = var_4192, x = query_states_61_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> var_4196 = const()[name = string("op_4196"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_4197_cast_fp16 = reshape(shape = var_4196, x = key_states_101_cast_fp16)[name = string("op_4197_cast_fp16")];
            tensor<int32, [4]> var_4203 = const()[name = string("op_4203"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_4204_cast_fp16 = reshape(shape = var_4203, x = value_states_61_cast_fp16)[name = string("op_4204_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_4208_cast_fp16 = mul(x = x_101_cast_fp16, y = var_859_cast_fp16)[name = string("op_4208_cast_fp16")];
            tensor<int32, [2]> var_4209_split_sizes_0 = const()[name = string("op_4209_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4209_axis_0 = const()[name = string("op_4209_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_4209_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_4209_cast_fp16_1 = split(axis = var_4209_axis_0, split_sizes = var_4209_split_sizes_0, x = x_101_cast_fp16)[name = string("op_4209_cast_fp16")];
            fp16 const_127_promoted_to_fp16 = const()[name = string("const_127_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_4211_cast_fp16 = mul(x = var_4209_cast_fp16_1, y = const_127_promoted_to_fp16)[name = string("op_4211_cast_fp16")];
            int32 var_4213 = const()[name = string("op_4213"), val = int32(-2)];
            bool var_4214_interleave_0 = const()[name = string("op_4214_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_4214_cast_fp16 = concat(axis = var_4213, interleave = var_4214_interleave_0, values = (var_4211_cast_fp16, var_4209_cast_fp16_0))[name = string("op_4214_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4215_cast_fp16 = mul(x = var_4214_cast_fp16, y = var_866_cast_fp16)[name = string("op_4215_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_63_cast_fp16 = add(x = var_4208_cast_fp16, y = var_4215_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4221_cast_fp16 = mul(x = var_4197_cast_fp16, y = var_859_cast_fp16)[name = string("op_4221_cast_fp16")];
            tensor<int32, [2]> var_4222_split_sizes_0 = const()[name = string("op_4222_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4222_axis_0 = const()[name = string("op_4222_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_4222_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_4222_cast_fp16_1 = split(axis = var_4222_axis_0, split_sizes = var_4222_split_sizes_0, x = var_4197_cast_fp16)[name = string("op_4222_cast_fp16")];
            fp16 const_128_promoted_to_fp16 = const()[name = string("const_128_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_4224_cast_fp16 = mul(x = var_4222_cast_fp16_1, y = const_128_promoted_to_fp16)[name = string("op_4224_cast_fp16")];
            int32 var_4226 = const()[name = string("op_4226"), val = int32(-2)];
            bool var_4227_interleave_0 = const()[name = string("op_4227_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_4227_cast_fp16 = concat(axis = var_4226, interleave = var_4227_interleave_0, values = (var_4224_cast_fp16, var_4222_cast_fp16_0))[name = string("op_4227_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4228_cast_fp16 = mul(x = var_4227_cast_fp16, y = var_866_cast_fp16)[name = string("op_4228_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_105_cast_fp16 = add(x = var_4221_cast_fp16, y = var_4228_cast_fp16)[name = string("key_states_105_cast_fp16")];
            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, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([10])];
            tensor<int32, [1]> expand_dims_122 = const()[name = string("expand_dims_122"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_124 = const()[name = string("expand_dims_124"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_125 = const()[name = string("expand_dims_125"), val = tensor<int32, [1]>([11])];
            int32 concat_83_axis_0 = const()[name = string("concat_83_axis_0"), val = int32(0)];
            bool concat_83_interleave_0 = const()[name = string("concat_83_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_83 = concat(axis = concat_83_axis_0, interleave = concat_83_interleave_0, values = (expand_dims_121, expand_dims_122, position_id, expand_dims_124))[name = string("concat_83")];
            tensor<int32, [1]> concat_84_values1_0 = const()[name = string("concat_84_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_84_values3_0 = const()[name = string("concat_84_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_84_axis_0 = const()[name = string("concat_84_axis_0"), val = int32(0)];
            bool concat_84_interleave_0 = const()[name = string("concat_84_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_84 = concat(axis = concat_84_axis_0, interleave = concat_84_interleave_0, values = (expand_dims_125, concat_84_values1_0, cache_position_end, concat_84_values3_0))[name = string("concat_84")];
            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_52")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_83, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_84, 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, [28, 2, 2048, 128]> coreml_update_state_76 = read_state(input = key_cache)[name = string("coreml_update_state_76")];
            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_4204_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_83, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_84, 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, [28, 2, 2048, 128]> coreml_update_state_77 = read_state(input = value_cache)[name = string("coreml_update_state_77")];
            tensor<int32, [4]> var_4298_begin_0 = const()[name = string("op_4298_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4298_end_0 = const()[name = string("op_4298_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_4298_end_mask_0 = const()[name = string("op_4298_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4298_cast_fp16 = slice_by_index(begin = var_4298_begin_0, end = var_4298_end_0, end_mask = var_4298_end_mask_0, x = coreml_update_state_76)[name = string("op_4298_cast_fp16")];
            tensor<int32, [2]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4301_axis_0 = const()[name = string("op_4301_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4301_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4301_cast_fp16_1 = split(axis = var_4301_axis_0, split_sizes = tile_20, x = var_4298_cast_fp16)[name = string("op_4301_cast_fp16")];
            tensor<int32, [4]> var_4308_begin_0 = const()[name = string("op_4308_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4308_end_0 = const()[name = string("op_4308_end_0"), val = tensor<int32, [4]>([11, 2, 2048, 128])];
            tensor<bool, [4]> var_4308_end_mask_0 = const()[name = string("op_4308_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4308_cast_fp16 = slice_by_index(begin = var_4308_begin_0, end = var_4308_end_0, end_mask = var_4308_end_mask_0, x = coreml_update_state_77)[name = string("op_4308_cast_fp16")];
            tensor<int32, [2]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4311_axis_0 = const()[name = string("op_4311_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4311_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4311_cast_fp16_1 = split(axis = var_4311_axis_0, split_sizes = tile_21, x = var_4308_cast_fp16)[name = string("op_4311_cast_fp16")];
            tensor<int32, [2]> var_4314_split_sizes_0 = const()[name = string("op_4314_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4314_axis_0 = const()[name = string("op_4314_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_4314_0, tensor<fp16, [1, 8, 128, 8]> var_4314_1 = split(axis = var_4314_axis_0, split_sizes = var_4314_split_sizes_0, x = query_states_63_cast_fp16)[name = string("op_4314")];
            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_4301_cast_fp16_0, y = var_4314_0)[name = string("attn_weights_161_cast_fp16")];
            fp16 var_4317_to_fp16 = const()[name = string("op_4317_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_4317_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_4321 = const()[name = string("op_4321"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_167_cast_fp16 = softmax(axis = var_4321, x = attn_weights_165_cast_fp16)[name = string("attn_weights_167_cast_fp16")];
            bool var_4327_transpose_x_1 = const()[name = string("op_4327_transpose_x_1"), val = bool(true)];
            bool var_4327_transpose_y_1 = const()[name = string("op_4327_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_4327_cast_fp16 = matmul(transpose_x = var_4327_transpose_x_1, transpose_y = var_4327_transpose_y_1, x = attn_weights_167_cast_fp16, y = var_4311_cast_fp16_0)[name = string("op_4327_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_4301_cast_fp16_1, y = var_4314_1)[name = string("attn_weights_169_cast_fp16")];
            fp16 var_4329_to_fp16 = const()[name = string("op_4329_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_4329_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_4333 = const()[name = string("op_4333"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_175_cast_fp16 = softmax(axis = var_4333, 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_4311_cast_fp16_1)[name = string("attn_output_81_cast_fp16")];
            int32 var_4341 = const()[name = string("op_4341"), 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_4341, interleave = attn_output_83_interleave_0, values = (var_4327_cast_fp16, attn_output_81_cast_fp16))[name = string("attn_output_83_cast_fp16")];
            tensor<int32, [4]> var_4345_perm_0 = const()[name = string("op_4345_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_4350 = const()[name = string("op_4350"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_4345_cast_fp16 = transpose(perm = var_4345_perm_0, x = attn_output_83_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_87_cast_fp16 = reshape(shape = var_4350, x = var_4345_cast_fp16)[name = string("attn_output_87_cast_fp16")];
            string hidden_states_103_pad_type_0 = const()[name = string("hidden_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_103_strides_0 = const()[name = string("hidden_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_4376 = const()[name = string("op_4376"), val = int32(1)];
            fp16 const_133_promoted_to_fp16 = const()[name = string("const_133_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4378_cast_fp16 = mul(x = hidden_states_105_cast_fp16, y = const_133_promoted_to_fp16)[name = string("op_4378_cast_fp16")];
            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_4376, interleave = doubled_85_interleave_0, values = (hidden_states_105_cast_fp16, var_4378_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(1407002304)))];
            fp16 var_4388_to_fp16 = const()[name = string("op_4388_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_4388_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [2]> var_4399_split_sizes_0 = const()[name = string("op_4399_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4399_axis_0 = const()[name = string("op_4399_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4399_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4399_cast_fp16_1 = split(axis = var_4399_axis_0, split_sizes = var_4399_split_sizes_0, x = out_43_cast_fp16)[name = string("op_4399_cast_fp16")];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_4399_cast_fp16_0)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_4416_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_4416_cast_fp16")];
            string var_4422_pad_type_0 = const()[name = string("op_4422_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4422_strides_0 = const()[name = string("op_4422_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4422_pad_0 = const()[name = string("op_4422_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4422_dilations_0 = const()[name = string("op_4422_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4422_groups_0 = const()[name = string("op_4422_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_4422_cast_fp16 = conv(dilations = var_4422_dilations_0, groups = var_4422_groups_0, pad = var_4422_pad_0, pad_type = var_4422_pad_type_0, strides = var_4422_strides_0, weight = layers_10_mlp_up_proj_weight_cast_fp16, x = var_4399_cast_fp16_0)[name = string("op_4422_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_109_cast_fp16 = mul(x = var_4416_cast_fp16, y = var_4422_cast_fp16)[name = string("x_109_cast_fp16")];
            string hidden_states_107_pad_type_0 = const()[name = string("hidden_states_107_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_107_strides_0 = const()[name = string("hidden_states_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_4438 = const()[name = string("op_4438"), val = int32(1)];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4440_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = const_135_promoted_to_fp16)[name = string("op_4440_cast_fp16")];
            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_4438, interleave = doubled_89_interleave_0, values = (hidden_states_109_cast_fp16, var_4440_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(1407010560)))];
            fp16 var_4450_to_fp16 = const()[name = string("op_4450_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_4450_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [2]> var_4461_split_sizes_0 = const()[name = string("op_4461_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4461_axis_0 = const()[name = string("op_4461_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4461_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4461_cast_fp16_1 = split(axis = var_4461_axis_0, split_sizes = var_4461_split_sizes_0, x = out_45_cast_fp16)[name = string("op_4461_cast_fp16")];
            string query_states_67_pad_type_0 = const()[name = string("query_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_67_strides_0 = const()[name = string("query_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_4461_cast_fp16_0)[name = string("query_states_67_cast_fp16")];
            string key_states_111_pad_type_0 = const()[name = string("key_states_111_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_111_strides_0 = const()[name = string("key_states_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_4461_cast_fp16_0)[name = string("key_states_111_cast_fp16")];
            string value_states_67_pad_type_0 = const()[name = string("value_states_67_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_67_strides_0 = const()[name = string("value_states_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_4461_cast_fp16_0)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> var_4513 = const()[name = string("op_4513"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_111_cast_fp16 = reshape(shape = var_4513, x = query_states_67_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [4]> var_4517 = const()[name = string("op_4517"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_4518_cast_fp16 = reshape(shape = var_4517, x = key_states_111_cast_fp16)[name = string("op_4518_cast_fp16")];
            tensor<int32, [4]> var_4524 = const()[name = string("op_4524"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_4525_cast_fp16 = reshape(shape = var_4524, x = value_states_67_cast_fp16)[name = string("op_4525_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_4529_cast_fp16 = mul(x = x_111_cast_fp16, y = var_859_cast_fp16)[name = string("op_4529_cast_fp16")];
            tensor<int32, [2]> var_4530_split_sizes_0 = const()[name = string("op_4530_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4530_axis_0 = const()[name = string("op_4530_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_4530_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_4530_cast_fp16_1 = split(axis = var_4530_axis_0, split_sizes = var_4530_split_sizes_0, x = x_111_cast_fp16)[name = string("op_4530_cast_fp16")];
            fp16 const_139_promoted_to_fp16 = const()[name = string("const_139_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_4532_cast_fp16 = mul(x = var_4530_cast_fp16_1, y = const_139_promoted_to_fp16)[name = string("op_4532_cast_fp16")];
            int32 var_4534 = const()[name = string("op_4534"), val = int32(-2)];
            bool var_4535_interleave_0 = const()[name = string("op_4535_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_4535_cast_fp16 = concat(axis = var_4534, interleave = var_4535_interleave_0, values = (var_4532_cast_fp16, var_4530_cast_fp16_0))[name = string("op_4535_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4536_cast_fp16 = mul(x = var_4535_cast_fp16, y = var_866_cast_fp16)[name = string("op_4536_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_69_cast_fp16 = add(x = var_4529_cast_fp16, y = var_4536_cast_fp16)[name = string("query_states_69_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4542_cast_fp16 = mul(x = var_4518_cast_fp16, y = var_859_cast_fp16)[name = string("op_4542_cast_fp16")];
            tensor<int32, [2]> var_4543_split_sizes_0 = const()[name = string("op_4543_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4543_axis_0 = const()[name = string("op_4543_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_4543_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_4543_cast_fp16_1 = split(axis = var_4543_axis_0, split_sizes = var_4543_split_sizes_0, x = var_4518_cast_fp16)[name = string("op_4543_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_4545_cast_fp16 = mul(x = var_4543_cast_fp16_1, y = const_140_promoted_to_fp16)[name = string("op_4545_cast_fp16")];
            int32 var_4547 = const()[name = string("op_4547"), val = int32(-2)];
            bool var_4548_interleave_0 = const()[name = string("op_4548_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_4548_cast_fp16 = concat(axis = var_4547, interleave = var_4548_interleave_0, values = (var_4545_cast_fp16, var_4543_cast_fp16_0))[name = string("op_4548_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4549_cast_fp16 = mul(x = var_4548_cast_fp16, y = var_866_cast_fp16)[name = string("op_4549_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_115_cast_fp16 = add(x = var_4542_cast_fp16, y = var_4549_cast_fp16)[name = string("key_states_115_cast_fp16")];
            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, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([11])];
            tensor<int32, [1]> expand_dims_134 = const()[name = string("expand_dims_134"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_136 = const()[name = string("expand_dims_136"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_137 = const()[name = string("expand_dims_137"), val = tensor<int32, [1]>([12])];
            int32 concat_91_axis_0 = const()[name = string("concat_91_axis_0"), val = int32(0)];
            bool concat_91_interleave_0 = const()[name = string("concat_91_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_91 = concat(axis = concat_91_axis_0, interleave = concat_91_interleave_0, values = (expand_dims_133, expand_dims_134, position_id, expand_dims_136))[name = string("concat_91")];
            tensor<int32, [1]> concat_92_values1_0 = const()[name = string("concat_92_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_92_values3_0 = const()[name = string("concat_92_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_92_axis_0 = const()[name = string("concat_92_axis_0"), val = int32(0)];
            bool concat_92_interleave_0 = const()[name = string("concat_92_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_92 = concat(axis = concat_92_axis_0, interleave = concat_92_interleave_0, values = (expand_dims_137, concat_92_values1_0, cache_position_end, concat_92_values3_0))[name = string("concat_92")];
            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_49")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_91, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_92, 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, [28, 2, 2048, 128]> coreml_update_state_78 = read_state(input = key_cache)[name = string("coreml_update_state_78")];
            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_4525_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_91, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_92, 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, [28, 2, 2048, 128]> coreml_update_state_79 = read_state(input = value_cache)[name = string("coreml_update_state_79")];
            tensor<int32, [4]> var_4619_begin_0 = const()[name = string("op_4619_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4619_end_0 = const()[name = string("op_4619_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4619_end_mask_0 = const()[name = string("op_4619_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4619_cast_fp16 = slice_by_index(begin = var_4619_begin_0, end = var_4619_end_0, end_mask = var_4619_end_mask_0, x = coreml_update_state_78)[name = string("op_4619_cast_fp16")];
            tensor<int32, [2]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4622_axis_0 = const()[name = string("op_4622_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4622_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4622_cast_fp16_1 = split(axis = var_4622_axis_0, split_sizes = tile_22, x = var_4619_cast_fp16)[name = string("op_4622_cast_fp16")];
            tensor<int32, [4]> var_4629_begin_0 = const()[name = string("op_4629_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4629_end_0 = const()[name = string("op_4629_end_0"), val = tensor<int32, [4]>([12, 2, 2048, 128])];
            tensor<bool, [4]> var_4629_end_mask_0 = const()[name = string("op_4629_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4629_cast_fp16 = slice_by_index(begin = var_4629_begin_0, end = var_4629_end_0, end_mask = var_4629_end_mask_0, x = coreml_update_state_79)[name = string("op_4629_cast_fp16")];
            tensor<int32, [2]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4632_axis_0 = const()[name = string("op_4632_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4632_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4632_cast_fp16_1 = split(axis = var_4632_axis_0, split_sizes = tile_23, x = var_4629_cast_fp16)[name = string("op_4632_cast_fp16")];
            tensor<int32, [2]> var_4635_split_sizes_0 = const()[name = string("op_4635_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4635_axis_0 = const()[name = string("op_4635_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_4635_0, tensor<fp16, [1, 8, 128, 8]> var_4635_1 = split(axis = var_4635_axis_0, split_sizes = var_4635_split_sizes_0, x = query_states_69_cast_fp16)[name = string("op_4635")];
            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_4622_cast_fp16_0, y = var_4635_0)[name = string("attn_weights_177_cast_fp16")];
            fp16 var_4638_to_fp16 = const()[name = string("op_4638_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_4638_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_4642 = const()[name = string("op_4642"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_183_cast_fp16 = softmax(axis = var_4642, x = attn_weights_181_cast_fp16)[name = string("attn_weights_183_cast_fp16")];
            bool var_4648_transpose_x_1 = const()[name = string("op_4648_transpose_x_1"), val = bool(true)];
            bool var_4648_transpose_y_1 = const()[name = string("op_4648_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_4648_cast_fp16 = matmul(transpose_x = var_4648_transpose_x_1, transpose_y = var_4648_transpose_y_1, x = attn_weights_183_cast_fp16, y = var_4632_cast_fp16_0)[name = string("op_4648_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_4622_cast_fp16_1, y = var_4635_1)[name = string("attn_weights_185_cast_fp16")];
            fp16 var_4650_to_fp16 = const()[name = string("op_4650_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_4650_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_4654 = const()[name = string("op_4654"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_191_cast_fp16 = softmax(axis = var_4654, 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_4632_cast_fp16_1)[name = string("attn_output_89_cast_fp16")];
            int32 var_4662 = const()[name = string("op_4662"), 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_4662, interleave = attn_output_91_interleave_0, values = (var_4648_cast_fp16, attn_output_89_cast_fp16))[name = string("attn_output_91_cast_fp16")];
            tensor<int32, [4]> var_4666_perm_0 = const()[name = string("op_4666_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_4671 = const()[name = string("op_4671"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_4666_cast_fp16 = transpose(perm = var_4666_perm_0, x = attn_output_91_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_95_cast_fp16 = reshape(shape = var_4671, x = var_4666_cast_fp16)[name = string("attn_output_95_cast_fp16")];
            string hidden_states_113_pad_type_0 = const()[name = string("hidden_states_113_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_113_strides_0 = const()[name = string("hidden_states_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_4697 = const()[name = string("op_4697"), val = int32(1)];
            fp16 const_145_promoted_to_fp16 = const()[name = string("const_145_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4699_cast_fp16 = mul(x = hidden_states_115_cast_fp16, y = const_145_promoted_to_fp16)[name = string("op_4699_cast_fp16")];
            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_4697, interleave = doubled_93_interleave_0, values = (hidden_states_115_cast_fp16, var_4699_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(1407018816)))];
            fp16 var_4709_to_fp16 = const()[name = string("op_4709_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_4709_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [2]> var_4720_split_sizes_0 = const()[name = string("op_4720_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4720_axis_0 = const()[name = string("op_4720_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4720_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4720_cast_fp16_1 = split(axis = var_4720_axis_0, split_sizes = var_4720_split_sizes_0, x = out_47_cast_fp16)[name = string("op_4720_cast_fp16")];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_4720_cast_fp16_0)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_4737_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_4737_cast_fp16")];
            string var_4743_pad_type_0 = const()[name = string("op_4743_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4743_strides_0 = const()[name = string("op_4743_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4743_pad_0 = const()[name = string("op_4743_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4743_dilations_0 = const()[name = string("op_4743_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4743_groups_0 = const()[name = string("op_4743_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_4743_cast_fp16 = conv(dilations = var_4743_dilations_0, groups = var_4743_groups_0, pad = var_4743_pad_0, pad_type = var_4743_pad_type_0, strides = var_4743_strides_0, weight = layers_11_mlp_up_proj_weight_cast_fp16, x = var_4720_cast_fp16_0)[name = string("op_4743_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_119_cast_fp16 = mul(x = var_4737_cast_fp16, y = var_4743_cast_fp16)[name = string("x_119_cast_fp16")];
            string hidden_states_117_pad_type_0 = const()[name = string("hidden_states_117_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_117_strides_0 = const()[name = string("hidden_states_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_4759 = const()[name = string("op_4759"), val = int32(1)];
            fp16 const_147_promoted_to_fp16 = const()[name = string("const_147_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_4761_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = const_147_promoted_to_fp16)[name = string("op_4761_cast_fp16")];
            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_4759, interleave = doubled_97_interleave_0, values = (hidden_states_119_cast_fp16, var_4761_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(1407027072)))];
            fp16 var_4771_to_fp16 = const()[name = string("op_4771_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_4771_to_fp16, gamma = out_49_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [2]> var_4782_split_sizes_0 = const()[name = string("op_4782_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_4782_axis_0 = const()[name = string("op_4782_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_4782_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_4782_cast_fp16_1 = split(axis = var_4782_axis_0, split_sizes = var_4782_split_sizes_0, x = out_49_cast_fp16)[name = string("op_4782_cast_fp16")];
            string query_states_73_pad_type_0 = const()[name = string("query_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_73_strides_0 = const()[name = string("query_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_4782_cast_fp16_0)[name = string("query_states_73_cast_fp16")];
            string key_states_121_pad_type_0 = const()[name = string("key_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_121_strides_0 = const()[name = string("key_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_4782_cast_fp16_0)[name = string("key_states_121_cast_fp16")];
            string value_states_73_pad_type_0 = const()[name = string("value_states_73_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_73_strides_0 = const()[name = string("value_states_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_4782_cast_fp16_0)[name = string("value_states_73_cast_fp16")];
            tensor<int32, [4]> var_4834 = const()[name = string("op_4834"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_121_cast_fp16 = reshape(shape = var_4834, x = query_states_73_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> var_4838 = const()[name = string("op_4838"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_4839_cast_fp16 = reshape(shape = var_4838, x = key_states_121_cast_fp16)[name = string("op_4839_cast_fp16")];
            tensor<int32, [4]> var_4845 = const()[name = string("op_4845"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_4846_cast_fp16 = reshape(shape = var_4845, x = value_states_73_cast_fp16)[name = string("op_4846_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_4850_cast_fp16 = mul(x = x_121_cast_fp16, y = var_859_cast_fp16)[name = string("op_4850_cast_fp16")];
            tensor<int32, [2]> var_4851_split_sizes_0 = const()[name = string("op_4851_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4851_axis_0 = const()[name = string("op_4851_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_4851_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_4851_cast_fp16_1 = split(axis = var_4851_axis_0, split_sizes = var_4851_split_sizes_0, x = x_121_cast_fp16)[name = string("op_4851_cast_fp16")];
            fp16 const_151_promoted_to_fp16 = const()[name = string("const_151_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_4853_cast_fp16 = mul(x = var_4851_cast_fp16_1, y = const_151_promoted_to_fp16)[name = string("op_4853_cast_fp16")];
            int32 var_4855 = const()[name = string("op_4855"), val = int32(-2)];
            bool var_4856_interleave_0 = const()[name = string("op_4856_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_4856_cast_fp16 = concat(axis = var_4855, interleave = var_4856_interleave_0, values = (var_4853_cast_fp16, var_4851_cast_fp16_0))[name = string("op_4856_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_4857_cast_fp16 = mul(x = var_4856_cast_fp16, y = var_866_cast_fp16)[name = string("op_4857_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_75_cast_fp16 = add(x = var_4850_cast_fp16, y = var_4857_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4863_cast_fp16 = mul(x = var_4839_cast_fp16, y = var_859_cast_fp16)[name = string("op_4863_cast_fp16")];
            tensor<int32, [2]> var_4864_split_sizes_0 = const()[name = string("op_4864_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4864_axis_0 = const()[name = string("op_4864_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_4864_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_4864_cast_fp16_1 = split(axis = var_4864_axis_0, split_sizes = var_4864_split_sizes_0, x = var_4839_cast_fp16)[name = string("op_4864_cast_fp16")];
            fp16 const_152_promoted_to_fp16 = const()[name = string("const_152_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_4866_cast_fp16 = mul(x = var_4864_cast_fp16_1, y = const_152_promoted_to_fp16)[name = string("op_4866_cast_fp16")];
            int32 var_4868 = const()[name = string("op_4868"), val = int32(-2)];
            bool var_4869_interleave_0 = const()[name = string("op_4869_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_4869_cast_fp16 = concat(axis = var_4868, interleave = var_4869_interleave_0, values = (var_4866_cast_fp16, var_4864_cast_fp16_0))[name = string("op_4869_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_4870_cast_fp16 = mul(x = var_4869_cast_fp16, y = var_866_cast_fp16)[name = string("op_4870_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_125_cast_fp16 = add(x = var_4863_cast_fp16, y = var_4870_cast_fp16)[name = string("key_states_125_cast_fp16")];
            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, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([12])];
            tensor<int32, [1]> expand_dims_146 = const()[name = string("expand_dims_146"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_148 = const()[name = string("expand_dims_148"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_149 = const()[name = string("expand_dims_149"), val = tensor<int32, [1]>([13])];
            int32 concat_99_axis_0 = const()[name = string("concat_99_axis_0"), val = int32(0)];
            bool concat_99_interleave_0 = const()[name = string("concat_99_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_99 = concat(axis = concat_99_axis_0, interleave = concat_99_interleave_0, values = (expand_dims_145, expand_dims_146, position_id, expand_dims_148))[name = string("concat_99")];
            tensor<int32, [1]> concat_100_values1_0 = const()[name = string("concat_100_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_100_values3_0 = const()[name = string("concat_100_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_100_axis_0 = const()[name = string("concat_100_axis_0"), val = int32(0)];
            bool concat_100_interleave_0 = const()[name = string("concat_100_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_100 = concat(axis = concat_100_axis_0, interleave = concat_100_interleave_0, values = (expand_dims_149, concat_100_values1_0, cache_position_end, concat_100_values3_0))[name = string("concat_100")];
            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_46")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_99, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_100, 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, [28, 2, 2048, 128]> coreml_update_state_80 = read_state(input = key_cache)[name = string("coreml_update_state_80")];
            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_4846_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_99, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_100, 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, [28, 2, 2048, 128]> coreml_update_state_81 = read_state(input = value_cache)[name = string("coreml_update_state_81")];
            tensor<int32, [4]> var_4940_begin_0 = const()[name = string("op_4940_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4940_end_0 = const()[name = string("op_4940_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4940_end_mask_0 = const()[name = string("op_4940_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4940_cast_fp16 = slice_by_index(begin = var_4940_begin_0, end = var_4940_end_0, end_mask = var_4940_end_mask_0, x = coreml_update_state_80)[name = string("op_4940_cast_fp16")];
            tensor<int32, [2]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4943_axis_0 = const()[name = string("op_4943_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4943_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4943_cast_fp16_1 = split(axis = var_4943_axis_0, split_sizes = tile_24, x = var_4940_cast_fp16)[name = string("op_4943_cast_fp16")];
            tensor<int32, [4]> var_4950_begin_0 = const()[name = string("op_4950_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_4950_end_0 = const()[name = string("op_4950_end_0"), val = tensor<int32, [4]>([13, 2, 2048, 128])];
            tensor<bool, [4]> var_4950_end_mask_0 = const()[name = string("op_4950_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_4950_cast_fp16 = slice_by_index(begin = var_4950_begin_0, end = var_4950_end_0, end_mask = var_4950_end_mask_0, x = coreml_update_state_81)[name = string("op_4950_cast_fp16")];
            tensor<int32, [2]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4953_axis_0 = const()[name = string("op_4953_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_4953_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_4953_cast_fp16_1 = split(axis = var_4953_axis_0, split_sizes = tile_25, x = var_4950_cast_fp16)[name = string("op_4953_cast_fp16")];
            tensor<int32, [2]> var_4956_split_sizes_0 = const()[name = string("op_4956_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_4956_axis_0 = const()[name = string("op_4956_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_4956_0, tensor<fp16, [1, 8, 128, 8]> var_4956_1 = split(axis = var_4956_axis_0, split_sizes = var_4956_split_sizes_0, x = query_states_75_cast_fp16)[name = string("op_4956")];
            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_4943_cast_fp16_0, y = var_4956_0)[name = string("attn_weights_193_cast_fp16")];
            fp16 var_4959_to_fp16 = const()[name = string("op_4959_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_4959_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_4963 = const()[name = string("op_4963"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_199_cast_fp16 = softmax(axis = var_4963, x = attn_weights_197_cast_fp16)[name = string("attn_weights_199_cast_fp16")];
            bool var_4969_transpose_x_1 = const()[name = string("op_4969_transpose_x_1"), val = bool(true)];
            bool var_4969_transpose_y_1 = const()[name = string("op_4969_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_4969_cast_fp16 = matmul(transpose_x = var_4969_transpose_x_1, transpose_y = var_4969_transpose_y_1, x = attn_weights_199_cast_fp16, y = var_4953_cast_fp16_0)[name = string("op_4969_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_4943_cast_fp16_1, y = var_4956_1)[name = string("attn_weights_201_cast_fp16")];
            fp16 var_4971_to_fp16 = const()[name = string("op_4971_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_4971_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_4975 = const()[name = string("op_4975"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_207_cast_fp16 = softmax(axis = var_4975, 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_4953_cast_fp16_1)[name = string("attn_output_97_cast_fp16")];
            int32 var_4983 = const()[name = string("op_4983"), 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_4983, interleave = attn_output_99_interleave_0, values = (var_4969_cast_fp16, attn_output_97_cast_fp16))[name = string("attn_output_99_cast_fp16")];
            tensor<int32, [4]> var_4987_perm_0 = const()[name = string("op_4987_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_4992 = const()[name = string("op_4992"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_4987_cast_fp16 = transpose(perm = var_4987_perm_0, x = attn_output_99_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_103_cast_fp16 = reshape(shape = var_4992, x = var_4987_cast_fp16)[name = string("attn_output_103_cast_fp16")];
            string hidden_states_123_pad_type_0 = const()[name = string("hidden_states_123_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_123_strides_0 = const()[name = string("hidden_states_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_5018 = const()[name = string("op_5018"), val = int32(1)];
            fp16 const_157_promoted_to_fp16 = const()[name = string("const_157_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_5020_cast_fp16 = mul(x = hidden_states_125_cast_fp16, y = const_157_promoted_to_fp16)[name = string("op_5020_cast_fp16")];
            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_5018, interleave = doubled_101_interleave_0, values = (hidden_states_125_cast_fp16, var_5020_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(1407035328)))];
            fp16 var_5030_to_fp16 = const()[name = string("op_5030_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_5030_to_fp16, gamma = out_51_gamma_0_to_fp16, x = doubled_101_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [2]> var_5041_split_sizes_0 = const()[name = string("op_5041_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_5041_axis_0 = const()[name = string("op_5041_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_5041_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_5041_cast_fp16_1 = split(axis = var_5041_axis_0, split_sizes = var_5041_split_sizes_0, x = out_51_cast_fp16)[name = string("op_5041_cast_fp16")];
            string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_5041_cast_fp16_0)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_5058_cast_fp16 = silu(x = input_25_cast_fp16)[name = string("op_5058_cast_fp16")];
            string var_5064_pad_type_0 = const()[name = string("op_5064_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5064_strides_0 = const()[name = string("op_5064_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5064_pad_0 = const()[name = string("op_5064_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5064_dilations_0 = const()[name = string("op_5064_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5064_groups_0 = const()[name = string("op_5064_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_5064_cast_fp16 = conv(dilations = var_5064_dilations_0, groups = var_5064_groups_0, pad = var_5064_pad_0, pad_type = var_5064_pad_type_0, strides = var_5064_strides_0, weight = layers_12_mlp_up_proj_weight_cast_fp16, x = var_5041_cast_fp16_0)[name = string("op_5064_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_129_cast_fp16 = mul(x = var_5058_cast_fp16, y = var_5064_cast_fp16)[name = string("x_129_cast_fp16")];
            string hidden_states_127_pad_type_0 = const()[name = string("hidden_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_127_strides_0 = const()[name = string("hidden_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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")];
            int32 var_5080 = const()[name = string("op_5080"), val = int32(1)];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_5082_cast_fp16 = mul(x = hidden_states_129_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_5082_cast_fp16")];
            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_5080, interleave = doubled_105_interleave_0, values = (hidden_states_129_cast_fp16, var_5082_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(1407043584)))];
            fp16 var_5092_to_fp16 = const()[name = string("op_5092_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_5092_to_fp16, gamma = out_53_gamma_0_to_fp16, x = doubled_105_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [2]> var_5103_split_sizes_0 = const()[name = string("op_5103_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_5103_axis_0 = const()[name = string("op_5103_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_5103_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_5103_cast_fp16_1 = split(axis = var_5103_axis_0, split_sizes = var_5103_split_sizes_0, x = out_53_cast_fp16)[name = string("op_5103_cast_fp16")];
            string query_states_79_pad_type_0 = const()[name = string("query_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_79_strides_0 = const()[name = string("query_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_5103_cast_fp16_0)[name = string("query_states_79_cast_fp16")];
            string key_states_131_pad_type_0 = const()[name = string("key_states_131_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_131_strides_0 = const()[name = string("key_states_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_5103_cast_fp16_0)[name = string("key_states_131_cast_fp16")];
            string value_states_79_pad_type_0 = const()[name = string("value_states_79_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_79_strides_0 = const()[name = string("value_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_5103_cast_fp16_0)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> var_5155 = const()[name = string("op_5155"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_131_cast_fp16 = reshape(shape = var_5155, x = query_states_79_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> var_5159 = const()[name = string("op_5159"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_5160_cast_fp16 = reshape(shape = var_5159, x = key_states_131_cast_fp16)[name = string("op_5160_cast_fp16")];
            tensor<int32, [4]> var_5166 = const()[name = string("op_5166"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_5167_cast_fp16 = reshape(shape = var_5166, x = value_states_79_cast_fp16)[name = string("op_5167_cast_fp16")];
            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<fp16, [1, 16, 128, 8]> var_5171_cast_fp16 = mul(x = x_131_cast_fp16, y = var_859_cast_fp16)[name = string("op_5171_cast_fp16")];
            tensor<int32, [2]> var_5172_split_sizes_0 = const()[name = string("op_5172_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5172_axis_0 = const()[name = string("op_5172_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_5172_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_5172_cast_fp16_1 = split(axis = var_5172_axis_0, split_sizes = var_5172_split_sizes_0, x = x_131_cast_fp16)[name = string("op_5172_cast_fp16")];
            fp16 const_163_promoted_to_fp16 = const()[name = string("const_163_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_5174_cast_fp16 = mul(x = var_5172_cast_fp16_1, y = const_163_promoted_to_fp16)[name = string("op_5174_cast_fp16")];
            int32 var_5176 = const()[name = string("op_5176"), val = int32(-2)];
            bool var_5177_interleave_0 = const()[name = string("op_5177_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_5177_cast_fp16 = concat(axis = var_5176, interleave = var_5177_interleave_0, values = (var_5174_cast_fp16, var_5172_cast_fp16_0))[name = string("op_5177_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_5178_cast_fp16 = mul(x = var_5177_cast_fp16, y = var_866_cast_fp16)[name = string("op_5178_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_81_cast_fp16 = add(x = var_5171_cast_fp16, y = var_5178_cast_fp16)[name = string("query_states_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_5184_cast_fp16 = mul(x = var_5160_cast_fp16, y = var_859_cast_fp16)[name = string("op_5184_cast_fp16")];
            tensor<int32, [2]> var_5185_split_sizes_0 = const()[name = string("op_5185_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5185_axis_0 = const()[name = string("op_5185_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_5185_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_5185_cast_fp16_1 = split(axis = var_5185_axis_0, split_sizes = var_5185_split_sizes_0, x = var_5160_cast_fp16)[name = string("op_5185_cast_fp16")];
            fp16 const_164_promoted_to_fp16 = const()[name = string("const_164_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_5187_cast_fp16 = mul(x = var_5185_cast_fp16_1, y = const_164_promoted_to_fp16)[name = string("op_5187_cast_fp16")];
            int32 var_5189 = const()[name = string("op_5189"), val = int32(-2)];
            bool var_5190_interleave_0 = const()[name = string("op_5190_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_5190_cast_fp16 = concat(axis = var_5189, interleave = var_5190_interleave_0, values = (var_5187_cast_fp16, var_5185_cast_fp16_0))[name = string("op_5190_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_5191_cast_fp16 = mul(x = var_5190_cast_fp16, y = var_866_cast_fp16)[name = string("op_5191_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_135_cast_fp16 = add(x = var_5184_cast_fp16, y = var_5191_cast_fp16)[name = string("key_states_135_cast_fp16")];
            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, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([13])];
            tensor<int32, [1]> expand_dims_158 = const()[name = string("expand_dims_158"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_160 = const()[name = string("expand_dims_160"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_161 = const()[name = string("expand_dims_161"), val = tensor<int32, [1]>([14])];
            int32 concat_107_axis_0 = const()[name = string("concat_107_axis_0"), val = int32(0)];
            bool concat_107_interleave_0 = const()[name = string("concat_107_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_107 = concat(axis = concat_107_axis_0, interleave = concat_107_interleave_0, values = (expand_dims_157, expand_dims_158, position_id, expand_dims_160))[name = string("concat_107")];
            tensor<int32, [1]> concat_108_values1_0 = const()[name = string("concat_108_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_108_values3_0 = const()[name = string("concat_108_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_108_axis_0 = const()[name = string("concat_108_axis_0"), val = int32(0)];
            bool concat_108_interleave_0 = const()[name = string("concat_108_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_108 = concat(axis = concat_108_axis_0, interleave = concat_108_interleave_0, values = (expand_dims_161, concat_108_values1_0, cache_position_end, concat_108_values3_0))[name = string("concat_108")];
            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_43")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_107, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_108, 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, [28, 2, 2048, 128]> coreml_update_state_82 = read_state(input = key_cache)[name = string("coreml_update_state_82")];
            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_5167_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_107, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_108, 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, [28, 2, 2048, 128]> coreml_update_state_83 = read_state(input = value_cache)[name = string("coreml_update_state_83")];
            tensor<int32, [4]> var_5261_begin_0 = const()[name = string("op_5261_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5261_end_0 = const()[name = string("op_5261_end_0"), val = tensor<int32, [4]>([14, 2, 2048, 128])];
            tensor<bool, [4]> var_5261_end_mask_0 = const()[name = string("op_5261_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_5261_cast_fp16 = slice_by_index(begin = var_5261_begin_0, end = var_5261_end_0, end_mask = var_5261_end_mask_0, x = coreml_update_state_82)[name = string("op_5261_cast_fp16")];
            tensor<int32, [2]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5264_axis_0 = const()[name = string("op_5264_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_5264_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_5264_cast_fp16_1 = split(axis = var_5264_axis_0, split_sizes = tile_26, x = var_5261_cast_fp16)[name = string("op_5264_cast_fp16")];
            tensor<int32, [4]> var_5271_begin_0 = const()[name = string("op_5271_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5271_end_0 = const()[name = string("op_5271_end_0"), val = tensor<int32, [4]>([14, 2, 2048, 128])];
            tensor<bool, [4]> var_5271_end_mask_0 = const()[name = string("op_5271_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_5271_cast_fp16 = slice_by_index(begin = var_5271_begin_0, end = var_5271_end_0, end_mask = var_5271_end_mask_0, x = coreml_update_state_83)[name = string("op_5271_cast_fp16")];
            tensor<int32, [2]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5274_axis_0 = const()[name = string("op_5274_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_5274_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_5274_cast_fp16_1 = split(axis = var_5274_axis_0, split_sizes = tile_27, x = var_5271_cast_fp16)[name = string("op_5274_cast_fp16")];
            tensor<int32, [2]> var_5277_split_sizes_0 = const()[name = string("op_5277_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_5277_axis_0 = const()[name = string("op_5277_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_5277_0, tensor<fp16, [1, 8, 128, 8]> var_5277_1 = split(axis = var_5277_axis_0, split_sizes = var_5277_split_sizes_0, x = query_states_81_cast_fp16)[name = string("op_5277")];
            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_5264_cast_fp16_0, y = var_5277_0)[name = string("attn_weights_209_cast_fp16")];
            fp16 var_5280_to_fp16 = const()[name = string("op_5280_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_5280_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_5284 = const()[name = string("op_5284"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_215_cast_fp16 = softmax(axis = var_5284, x = attn_weights_213_cast_fp16)[name = string("attn_weights_215_cast_fp16")];
            bool var_5290_transpose_x_1 = const()[name = string("op_5290_transpose_x_1"), val = bool(true)];
            bool var_5290_transpose_y_1 = const()[name = string("op_5290_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_5290_cast_fp16 = matmul(transpose_x = var_5290_transpose_x_1, transpose_y = var_5290_transpose_y_1, x = attn_weights_215_cast_fp16, y = var_5274_cast_fp16_0)[name = string("op_5290_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_5264_cast_fp16_1, y = var_5277_1)[name = string("attn_weights_217_cast_fp16")];
            fp16 var_5292_to_fp16 = const()[name = string("op_5292_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_5292_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_5296 = const()[name = string("op_5296"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_223_cast_fp16 = softmax(axis = var_5296, x = attn_weights_221_cast_fp16)[name = string("attn_weights_223_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_223_cast_fp16, y = var_5274_cast_fp16_1)[name = string("attn_output_105_cast_fp16")];
            int32 var_5304 = const()[name = string("op_5304"), 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_5304, interleave = attn_output_107_interleave_0, values = (var_5290_cast_fp16, attn_output_105_cast_fp16))[name = string("attn_output_107_cast_fp16")];
            tensor<int32, [4]> var_5308_perm_0 = const()[name = string("op_5308_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_5313 = const()[name = string("op_5313"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_5308_cast_fp16 = transpose(perm = var_5308_perm_0, x = attn_output_107_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_111_cast_fp16 = reshape(shape = var_5313, x = var_5308_cast_fp16)[name = string("attn_output_111_cast_fp16")];
            string hidden_states_133_pad_type_0 = const()[name = string("hidden_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_133_strides_0 = const()[name = string("hidden_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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_111_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")];
            int32 var_5339 = const()[name = string("op_5339"), val = int32(1)];
            fp16 const_169_promoted_to_fp16 = const()[name = string("const_169_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_5341_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = const_169_promoted_to_fp16)[name = string("op_5341_cast_fp16")];
            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_5339, interleave = doubled_109_interleave_0, values = (hidden_states_135_cast_fp16, var_5341_cast_fp16))[name = string("doubled_109_cast_fp16")];
            tensor<int32, [1]> out_55_axes_0 = const()[name = string("out_55_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_55_gamma_0_to_fp16 = const()[name = string("out_55_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1407051840)))];
            fp16 var_5351_to_fp16 = const()[name = string("op_5351_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_5351_to_fp16, gamma = out_55_gamma_0_to_fp16, x = doubled_109_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<int32, [2]> var_5362_split_sizes_0 = const()[name = string("op_5362_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_5362_axis_0 = const()[name = string("op_5362_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_5362_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_5362_cast_fp16_1 = split(axis = var_5362_axis_0, split_sizes = var_5362_split_sizes_0, x = out_55_cast_fp16)[name = string("op_5362_cast_fp16")];
            string input_27_pad_type_0 = const()[name = string("input_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_27_strides_0 = const()[name = string("input_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_27_pad_0 = const()[name = string("input_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_27_dilations_0 = const()[name = string("input_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_27_groups_0 = const()[name = string("input_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_27_cast_fp16 = conv(dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = layers_13_mlp_gate_proj_weight_cast_fp16, x = var_5362_cast_fp16_0)[name = string("input_27_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_5379_cast_fp16 = silu(x = input_27_cast_fp16)[name = string("op_5379_cast_fp16")];
            string var_5385_pad_type_0 = const()[name = string("op_5385_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5385_strides_0 = const()[name = string("op_5385_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5385_pad_0 = const()[name = string("op_5385_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5385_dilations_0 = const()[name = string("op_5385_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5385_groups_0 = const()[name = string("op_5385_groups_0"), val = int32(1)];
            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(1407060096)))];
            tensor<fp16, [1, 6144, 1, 8]> var_5385_cast_fp16 = conv(dilations = var_5385_dilations_0, groups = var_5385_groups_0, pad = var_5385_pad_0, pad_type = var_5385_pad_type_0, strides = var_5385_strides_0, weight = layers_13_mlp_up_proj_weight_to_fp16, x = var_5362_cast_fp16_0)[name = string("op_5385_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_139_cast_fp16 = mul(x = var_5379_cast_fp16, y = var_5385_cast_fp16)[name = string("x_139_cast_fp16")];
            string hidden_states_137_pad_type_0 = const()[name = string("hidden_states_137_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_137_strides_0 = const()[name = string("hidden_states_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_137_pad_0 = const()[name = string("hidden_states_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_137_dilations_0 = const()[name = string("hidden_states_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_137_groups_0 = const()[name = string("hidden_states_137_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_137_cast_fp16 = conv(dilations = hidden_states_137_dilations_0, groups = hidden_states_137_groups_0, pad = hidden_states_137_pad_0, pad_type = hidden_states_137_pad_type_0, strides = hidden_states_137_strides_0, weight = layers_13_mlp_down_proj_weight_cast_fp16, x = x_139_cast_fp16)[name = string("hidden_states_137_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_139_cast_fp16 = add(x = hidden_states_135_cast_fp16, y = hidden_states_137_cast_fp16)[name = string("hidden_states_139_cast_fp16")];
            int32 var_5401 = const()[name = string("op_5401"), val = int32(1)];
            fp16 const_171_promoted_to_fp16 = const()[name = string("const_171_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_5403_cast_fp16 = mul(x = hidden_states_139_cast_fp16, y = const_171_promoted_to_fp16)[name = string("op_5403_cast_fp16")];
            bool doubled_113_interleave_0 = const()[name = string("doubled_113_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_113_cast_fp16 = concat(axis = var_5401, interleave = doubled_113_interleave_0, values = (hidden_states_139_cast_fp16, var_5403_cast_fp16))[name = string("doubled_113_cast_fp16")];
            tensor<int32, [1]> out_57_axes_0 = const()[name = string("out_57_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_57_gamma_0_to_fp16 = const()[name = string("out_57_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1432225984)))];
            fp16 var_5413_to_fp16 = const()[name = string("op_5413_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_5413_to_fp16, gamma = out_57_gamma_0_to_fp16, x = doubled_113_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<int32, [2]> var_5424_split_sizes_0 = const()[name = string("op_5424_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_5424_axis_0 = const()[name = string("op_5424_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_5424_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_5424_cast_fp16_1 = split(axis = var_5424_axis_0, split_sizes = var_5424_split_sizes_0, x = out_57_cast_fp16)[name = string("op_5424_cast_fp16")];
            string query_states_85_pad_type_0 = const()[name = string("query_states_85_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_85_strides_0 = const()[name = string("query_states_85_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_85_pad_0 = const()[name = string("query_states_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_85_dilations_0 = const()[name = string("query_states_85_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_85_groups_0 = const()[name = string("query_states_85_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_85_cast_fp16 = conv(dilations = query_states_85_dilations_0, groups = query_states_85_groups_0, pad = query_states_85_pad_0, pad_type = query_states_85_pad_type_0, strides = query_states_85_strides_0, weight = layers_14_self_attn_q_proj_weight_cast_fp16, x = var_5424_cast_fp16_0)[name = string("query_states_85_cast_fp16")];
            string key_states_141_pad_type_0 = const()[name = string("key_states_141_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_141_strides_0 = const()[name = string("key_states_141_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_141_pad_0 = const()[name = string("key_states_141_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_141_dilations_0 = const()[name = string("key_states_141_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_141_groups_0 = const()[name = string("key_states_141_groups_0"), val = int32(1)];
            tensor<fp16, [256, 2048, 1, 1]> layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1432234240)))];
            tensor<fp16, [1, 256, 1, 8]> key_states_141_cast_fp16 = conv(dilations = key_states_141_dilations_0, groups = key_states_141_groups_0, pad = key_states_141_pad_0, pad_type = key_states_141_pad_type_0, strides = key_states_141_strides_0, weight = layers_14_self_attn_k_proj_weight_to_fp16, x = var_5424_cast_fp16_0)[name = string("key_states_141_cast_fp16")];
            string value_states_85_pad_type_0 = const()[name = string("value_states_85_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_85_strides_0 = const()[name = string("value_states_85_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_85_pad_0 = const()[name = string("value_states_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_85_dilations_0 = const()[name = string("value_states_85_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_85_groups_0 = const()[name = string("value_states_85_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_85_cast_fp16 = conv(dilations = value_states_85_dilations_0, groups = value_states_85_groups_0, pad = value_states_85_pad_0, pad_type = value_states_85_pad_type_0, strides = value_states_85_strides_0, weight = layers_14_self_attn_v_proj_weight_cast_fp16, x = var_5424_cast_fp16_0)[name = string("value_states_85_cast_fp16")];
            tensor<int32, [4]> var_5476 = const()[name = string("op_5476"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_141_cast_fp16 = reshape(shape = var_5476, x = query_states_85_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<int32, [4]> var_5480 = const()[name = string("op_5480"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_5481_cast_fp16 = reshape(shape = var_5480, x = key_states_141_cast_fp16)[name = string("op_5481_cast_fp16")];
            tensor<int32, [4]> var_5487 = const()[name = string("op_5487"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_5488_cast_fp16 = reshape(shape = var_5487, x = value_states_85_cast_fp16)[name = string("op_5488_cast_fp16")];
            tensor<int32, [4]> value_states_87_perm_0 = const()[name = string("value_states_87_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_5492_cast_fp16 = mul(x = x_141_cast_fp16, y = var_859_cast_fp16)[name = string("op_5492_cast_fp16")];
            tensor<int32, [2]> var_5493_split_sizes_0 = const()[name = string("op_5493_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5493_axis_0 = const()[name = string("op_5493_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_5493_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_5493_cast_fp16_1 = split(axis = var_5493_axis_0, split_sizes = var_5493_split_sizes_0, x = x_141_cast_fp16)[name = string("op_5493_cast_fp16")];
            fp16 const_175_promoted_to_fp16 = const()[name = string("const_175_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_5495_cast_fp16 = mul(x = var_5493_cast_fp16_1, y = const_175_promoted_to_fp16)[name = string("op_5495_cast_fp16")];
            int32 var_5497 = const()[name = string("op_5497"), val = int32(-2)];
            bool var_5498_interleave_0 = const()[name = string("op_5498_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_5498_cast_fp16 = concat(axis = var_5497, interleave = var_5498_interleave_0, values = (var_5495_cast_fp16, var_5493_cast_fp16_0))[name = string("op_5498_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_5499_cast_fp16 = mul(x = var_5498_cast_fp16, y = var_866_cast_fp16)[name = string("op_5499_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_87_cast_fp16 = add(x = var_5492_cast_fp16, y = var_5499_cast_fp16)[name = string("query_states_87_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_5505_cast_fp16 = mul(x = var_5481_cast_fp16, y = var_859_cast_fp16)[name = string("op_5505_cast_fp16")];
            tensor<int32, [2]> var_5506_split_sizes_0 = const()[name = string("op_5506_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5506_axis_0 = const()[name = string("op_5506_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_5506_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_5506_cast_fp16_1 = split(axis = var_5506_axis_0, split_sizes = var_5506_split_sizes_0, x = var_5481_cast_fp16)[name = string("op_5506_cast_fp16")];
            fp16 const_176_promoted_to_fp16 = const()[name = string("const_176_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_5508_cast_fp16 = mul(x = var_5506_cast_fp16_1, y = const_176_promoted_to_fp16)[name = string("op_5508_cast_fp16")];
            int32 var_5510 = const()[name = string("op_5510"), val = int32(-2)];
            bool var_5511_interleave_0 = const()[name = string("op_5511_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_5511_cast_fp16 = concat(axis = var_5510, interleave = var_5511_interleave_0, values = (var_5508_cast_fp16, var_5506_cast_fp16_0))[name = string("op_5511_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_5512_cast_fp16 = mul(x = var_5511_cast_fp16, y = var_866_cast_fp16)[name = string("op_5512_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_145_cast_fp16 = add(x = var_5505_cast_fp16, y = var_5512_cast_fp16)[name = string("key_states_145_cast_fp16")];
            tensor<int32, [4]> key_states_147_perm_0 = const()[name = string("key_states_147_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_169 = const()[name = string("expand_dims_169"), val = tensor<int32, [1]>([14])];
            tensor<int32, [1]> expand_dims_170 = const()[name = string("expand_dims_170"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_172 = const()[name = string("expand_dims_172"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_173 = const()[name = string("expand_dims_173"), val = tensor<int32, [1]>([15])];
            int32 concat_115_axis_0 = const()[name = string("concat_115_axis_0"), val = int32(0)];
            bool concat_115_interleave_0 = const()[name = string("concat_115_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_115 = concat(axis = concat_115_axis_0, interleave = concat_115_interleave_0, values = (expand_dims_169, expand_dims_170, position_id, expand_dims_172))[name = string("concat_115")];
            tensor<int32, [1]> concat_116_values1_0 = const()[name = string("concat_116_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_116_values3_0 = const()[name = string("concat_116_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_116_axis_0 = const()[name = string("concat_116_axis_0"), val = int32(0)];
            bool concat_116_interleave_0 = const()[name = string("concat_116_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_116 = concat(axis = concat_116_axis_0, interleave = concat_116_interleave_0, values = (expand_dims_173, concat_116_values1_0, cache_position_end, concat_116_values3_0))[name = string("concat_116")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_15_stride_0 = const()[name = string("key_cache_internal_tensor_assign_15_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_15_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_15_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_15_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_15_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_15_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_15_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_147_cast_fp16 = transpose(perm = key_states_147_perm_0, x = key_states_145_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_15_cast_fp16 = slice_update(begin = concat_115, begin_mask = key_cache_internal_tensor_assign_15_begin_mask_0, end = concat_116, end_mask = key_cache_internal_tensor_assign_15_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_15_squeeze_mask_0, stride = key_cache_internal_tensor_assign_15_stride_0, update = key_states_147_cast_fp16, x = coreml_update_state_82)[name = string("key_cache_internal_tensor_assign_15_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_15_cast_fp16, input = key_cache)[name = string("coreml_update_state_84_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_84 = read_state(input = key_cache)[name = string("coreml_update_state_84")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_15_stride_0 = const()[name = string("value_cache_internal_tensor_assign_15_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_15_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_15_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_15_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_15_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_15_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_15_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_87_cast_fp16 = transpose(perm = value_states_87_perm_0, x = var_5488_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_15_cast_fp16 = slice_update(begin = concat_115, begin_mask = value_cache_internal_tensor_assign_15_begin_mask_0, end = concat_116, end_mask = value_cache_internal_tensor_assign_15_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_15_squeeze_mask_0, stride = value_cache_internal_tensor_assign_15_stride_0, update = value_states_87_cast_fp16, x = coreml_update_state_83)[name = string("value_cache_internal_tensor_assign_15_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_15_cast_fp16, input = value_cache)[name = string("coreml_update_state_85_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_85 = read_state(input = value_cache)[name = string("coreml_update_state_85")];
            tensor<int32, [4]> var_5582_begin_0 = const()[name = string("op_5582_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_5582_end_0 = const()[name = string("op_5582_end_0"), val = tensor<int32, [4]>([15, 2, 2048, 128])];
            tensor<bool, [4]> var_5582_end_mask_0 = const()[name = string("op_5582_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_5582_cast_fp16 = slice_by_index(begin = var_5582_begin_0, end = var_5582_end_0, end_mask = var_5582_end_mask_0, x = coreml_update_state_84)[name = string("op_5582_cast_fp16")];
            tensor<int32, [2]> tile_28 = const()[name = string("tile_28"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5585_axis_0 = const()[name = string("op_5585_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_5585_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_5585_cast_fp16_1 = split(axis = var_5585_axis_0, split_sizes = tile_28, x = var_5582_cast_fp16)[name = string("op_5585_cast_fp16")];
            tensor<int32, [4]> var_5592_begin_0 = const()[name = string("op_5592_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_5592_end_0 = const()[name = string("op_5592_end_0"), val = tensor<int32, [4]>([15, 2, 2048, 128])];
            tensor<bool, [4]> var_5592_end_mask_0 = const()[name = string("op_5592_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_5592_cast_fp16 = slice_by_index(begin = var_5592_begin_0, end = var_5592_end_0, end_mask = var_5592_end_mask_0, x = coreml_update_state_85)[name = string("op_5592_cast_fp16")];
            tensor<int32, [2]> tile_29 = const()[name = string("tile_29"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5595_axis_0 = const()[name = string("op_5595_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_5595_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_5595_cast_fp16_1 = split(axis = var_5595_axis_0, split_sizes = tile_29, x = var_5592_cast_fp16)[name = string("op_5595_cast_fp16")];
            tensor<int32, [2]> var_5598_split_sizes_0 = const()[name = string("op_5598_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_5598_axis_0 = const()[name = string("op_5598_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_5598_0, tensor<fp16, [1, 8, 128, 8]> var_5598_1 = split(axis = var_5598_axis_0, split_sizes = var_5598_split_sizes_0, x = query_states_87_cast_fp16)[name = string("op_5598")];
            bool attn_weights_225_transpose_x_0 = const()[name = string("attn_weights_225_transpose_x_0"), val = bool(false)];
            bool attn_weights_225_transpose_y_0 = const()[name = string("attn_weights_225_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_225_cast_fp16 = matmul(transpose_x = attn_weights_225_transpose_x_0, transpose_y = attn_weights_225_transpose_y_0, x = var_5585_cast_fp16_0, y = var_5598_0)[name = string("attn_weights_225_cast_fp16")];
            fp16 var_5601_to_fp16 = const()[name = string("op_5601_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_227_cast_fp16 = mul(x = attn_weights_225_cast_fp16, y = var_5601_to_fp16)[name = string("attn_weights_227_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_229_cast_fp16 = add(x = attn_weights_227_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_229_cast_fp16")];
            int32 var_5605 = const()[name = string("op_5605"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_231_cast_fp16 = softmax(axis = var_5605, x = attn_weights_229_cast_fp16)[name = string("attn_weights_231_cast_fp16")];
            bool var_5611_transpose_x_1 = const()[name = string("op_5611_transpose_x_1"), val = bool(true)];
            bool var_5611_transpose_y_1 = const()[name = string("op_5611_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_5611_cast_fp16 = matmul(transpose_x = var_5611_transpose_x_1, transpose_y = var_5611_transpose_y_1, x = attn_weights_231_cast_fp16, y = var_5595_cast_fp16_0)[name = string("op_5611_cast_fp16")];
            bool attn_weights_233_transpose_x_0 = const()[name = string("attn_weights_233_transpose_x_0"), val = bool(false)];
            bool attn_weights_233_transpose_y_0 = const()[name = string("attn_weights_233_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_233_cast_fp16 = matmul(transpose_x = attn_weights_233_transpose_x_0, transpose_y = attn_weights_233_transpose_y_0, x = var_5585_cast_fp16_1, y = var_5598_1)[name = string("attn_weights_233_cast_fp16")];
            fp16 var_5613_to_fp16 = const()[name = string("op_5613_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_235_cast_fp16 = mul(x = attn_weights_233_cast_fp16, y = var_5613_to_fp16)[name = string("attn_weights_235_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_237_cast_fp16 = add(x = attn_weights_235_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_237_cast_fp16")];
            int32 var_5617 = const()[name = string("op_5617"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_239_cast_fp16 = softmax(axis = var_5617, x = attn_weights_237_cast_fp16)[name = string("attn_weights_239_cast_fp16")];
            bool attn_output_113_transpose_x_1 = const()[name = string("attn_output_113_transpose_x_1"), val = bool(true)];
            bool attn_output_113_transpose_y_1 = const()[name = string("attn_output_113_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_113_cast_fp16 = matmul(transpose_x = attn_output_113_transpose_x_1, transpose_y = attn_output_113_transpose_y_1, x = attn_weights_239_cast_fp16, y = var_5595_cast_fp16_1)[name = string("attn_output_113_cast_fp16")];
            int32 var_5625 = const()[name = string("op_5625"), val = int32(1)];
            bool attn_output_115_interleave_0 = const()[name = string("attn_output_115_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_115_cast_fp16 = concat(axis = var_5625, interleave = attn_output_115_interleave_0, values = (var_5611_cast_fp16, attn_output_113_cast_fp16))[name = string("attn_output_115_cast_fp16")];
            tensor<int32, [4]> var_5629_perm_0 = const()[name = string("op_5629_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_5634 = const()[name = string("op_5634"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_5629_cast_fp16 = transpose(perm = var_5629_perm_0, x = attn_output_115_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_119_cast_fp16 = reshape(shape = var_5634, x = var_5629_cast_fp16)[name = string("attn_output_119_cast_fp16")];
            string hidden_states_143_pad_type_0 = const()[name = string("hidden_states_143_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_143_strides_0 = const()[name = string("hidden_states_143_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_143_pad_0 = const()[name = string("hidden_states_143_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_143_dilations_0 = const()[name = string("hidden_states_143_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_143_groups_0 = const()[name = string("hidden_states_143_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_143_cast_fp16 = conv(dilations = hidden_states_143_dilations_0, groups = hidden_states_143_groups_0, pad = hidden_states_143_pad_0, pad_type = hidden_states_143_pad_type_0, strides = hidden_states_143_strides_0, weight = layers_14_self_attn_o_proj_weight_cast_fp16, x = attn_output_119_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_145_cast_fp16 = add(x = hidden_states_139_cast_fp16, y = hidden_states_143_cast_fp16)[name = string("hidden_states_145_cast_fp16")];
            int32 var_5660 = const()[name = string("op_5660"), val = int32(1)];
            fp16 const_181_promoted_to_fp16 = const()[name = string("const_181_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_5662_cast_fp16 = mul(x = hidden_states_145_cast_fp16, y = const_181_promoted_to_fp16)[name = string("op_5662_cast_fp16")];
            bool doubled_117_interleave_0 = const()[name = string("doubled_117_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_117_cast_fp16 = concat(axis = var_5660, interleave = doubled_117_interleave_0, values = (hidden_states_145_cast_fp16, var_5662_cast_fp16))[name = string("doubled_117_cast_fp16")];
            tensor<int32, [1]> out_59_axes_0 = const()[name = string("out_59_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_59_gamma_0_to_fp16 = const()[name = string("out_59_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1433282880)))];
            fp16 var_5672_to_fp16 = const()[name = string("op_5672_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_5672_to_fp16, gamma = out_59_gamma_0_to_fp16, x = doubled_117_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<int32, [2]> var_5683_split_sizes_0 = const()[name = string("op_5683_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_5683_axis_0 = const()[name = string("op_5683_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_5683_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_5683_cast_fp16_1 = split(axis = var_5683_axis_0, split_sizes = var_5683_split_sizes_0, x = out_59_cast_fp16)[name = string("op_5683_cast_fp16")];
            string input_29_pad_type_0 = const()[name = string("input_29_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_29_strides_0 = const()[name = string("input_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_29_pad_0 = const()[name = string("input_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_29_dilations_0 = const()[name = string("input_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_29_groups_0 = const()[name = string("input_29_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_29_cast_fp16 = conv(dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = layers_14_mlp_gate_proj_weight_cast_fp16, x = var_5683_cast_fp16_0)[name = string("input_29_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_5700_cast_fp16 = silu(x = input_29_cast_fp16)[name = string("op_5700_cast_fp16")];
            string var_5706_pad_type_0 = const()[name = string("op_5706_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5706_strides_0 = const()[name = string("op_5706_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5706_pad_0 = const()[name = string("op_5706_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5706_dilations_0 = const()[name = string("op_5706_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5706_groups_0 = const()[name = string("op_5706_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_5706_cast_fp16 = conv(dilations = var_5706_dilations_0, groups = var_5706_groups_0, pad = var_5706_pad_0, pad_type = var_5706_pad_type_0, strides = var_5706_strides_0, weight = layers_14_mlp_up_proj_weight_cast_fp16, x = var_5683_cast_fp16_0)[name = string("op_5706_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_149_cast_fp16 = mul(x = var_5700_cast_fp16, y = var_5706_cast_fp16)[name = string("x_149_cast_fp16")];
            string hidden_states_147_pad_type_0 = const()[name = string("hidden_states_147_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_147_strides_0 = const()[name = string("hidden_states_147_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_147_pad_0 = const()[name = string("hidden_states_147_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_147_dilations_0 = const()[name = string("hidden_states_147_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_147_groups_0 = const()[name = string("hidden_states_147_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_147_cast_fp16 = conv(dilations = hidden_states_147_dilations_0, groups = hidden_states_147_groups_0, pad = hidden_states_147_pad_0, pad_type = hidden_states_147_pad_type_0, strides = hidden_states_147_strides_0, weight = layers_14_mlp_down_proj_weight_cast_fp16, x = x_149_cast_fp16)[name = string("hidden_states_147_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_149_cast_fp16 = add(x = hidden_states_145_cast_fp16, y = hidden_states_147_cast_fp16)[name = string("hidden_states_149_cast_fp16")];
            int32 var_5722 = const()[name = string("op_5722"), val = int32(1)];
            fp16 const_183_promoted_to_fp16 = const()[name = string("const_183_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_5724_cast_fp16 = mul(x = hidden_states_149_cast_fp16, y = const_183_promoted_to_fp16)[name = string("op_5724_cast_fp16")];
            bool doubled_121_interleave_0 = const()[name = string("doubled_121_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_121_cast_fp16 = concat(axis = var_5722, interleave = doubled_121_interleave_0, values = (hidden_states_149_cast_fp16, var_5724_cast_fp16))[name = string("doubled_121_cast_fp16")];
            tensor<int32, [1]> out_61_axes_0 = const()[name = string("out_61_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_61_gamma_0_to_fp16 = const()[name = string("out_61_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1433291136)))];
            fp16 var_5734_to_fp16 = const()[name = string("op_5734_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_5734_to_fp16, gamma = out_61_gamma_0_to_fp16, x = doubled_121_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<int32, [2]> var_5745_split_sizes_0 = const()[name = string("op_5745_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_5745_axis_0 = const()[name = string("op_5745_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_5745_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_5745_cast_fp16_1 = split(axis = var_5745_axis_0, split_sizes = var_5745_split_sizes_0, x = out_61_cast_fp16)[name = string("op_5745_cast_fp16")];
            string query_states_91_pad_type_0 = const()[name = string("query_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_91_strides_0 = const()[name = string("query_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_91_pad_0 = const()[name = string("query_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_91_dilations_0 = const()[name = string("query_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_91_groups_0 = const()[name = string("query_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_91_cast_fp16 = conv(dilations = query_states_91_dilations_0, groups = query_states_91_groups_0, pad = query_states_91_pad_0, pad_type = query_states_91_pad_type_0, strides = query_states_91_strides_0, weight = layers_15_self_attn_q_proj_weight_cast_fp16, x = var_5745_cast_fp16_0)[name = string("query_states_91_cast_fp16")];
            string key_states_151_pad_type_0 = const()[name = string("key_states_151_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_151_strides_0 = const()[name = string("key_states_151_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_151_pad_0 = const()[name = string("key_states_151_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_151_dilations_0 = const()[name = string("key_states_151_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_151_groups_0 = const()[name = string("key_states_151_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_151_cast_fp16 = conv(dilations = key_states_151_dilations_0, groups = key_states_151_groups_0, pad = key_states_151_pad_0, pad_type = key_states_151_pad_type_0, strides = key_states_151_strides_0, weight = layers_15_self_attn_k_proj_weight_cast_fp16, x = var_5745_cast_fp16_0)[name = string("key_states_151_cast_fp16")];
            string value_states_91_pad_type_0 = const()[name = string("value_states_91_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_91_strides_0 = const()[name = string("value_states_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_91_pad_0 = const()[name = string("value_states_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_91_dilations_0 = const()[name = string("value_states_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_91_groups_0 = const()[name = string("value_states_91_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_91_cast_fp16 = conv(dilations = value_states_91_dilations_0, groups = value_states_91_groups_0, pad = value_states_91_pad_0, pad_type = value_states_91_pad_type_0, strides = value_states_91_strides_0, weight = layers_15_self_attn_v_proj_weight_cast_fp16, x = var_5745_cast_fp16_0)[name = string("value_states_91_cast_fp16")];
            tensor<int32, [4]> var_5797 = const()[name = string("op_5797"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_151_cast_fp16 = reshape(shape = var_5797, x = query_states_91_cast_fp16)[name = string("x_151_cast_fp16")];
            tensor<int32, [4]> var_5801 = const()[name = string("op_5801"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_5802_cast_fp16 = reshape(shape = var_5801, x = key_states_151_cast_fp16)[name = string("op_5802_cast_fp16")];
            tensor<int32, [4]> var_5808 = const()[name = string("op_5808"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_5809_cast_fp16 = reshape(shape = var_5808, x = value_states_91_cast_fp16)[name = string("op_5809_cast_fp16")];
            tensor<int32, [4]> value_states_93_perm_0 = const()[name = string("value_states_93_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_5813_cast_fp16 = mul(x = x_151_cast_fp16, y = var_859_cast_fp16)[name = string("op_5813_cast_fp16")];
            tensor<int32, [2]> var_5814_split_sizes_0 = const()[name = string("op_5814_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5814_axis_0 = const()[name = string("op_5814_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_5814_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_5814_cast_fp16_1 = split(axis = var_5814_axis_0, split_sizes = var_5814_split_sizes_0, x = x_151_cast_fp16)[name = string("op_5814_cast_fp16")];
            fp16 const_187_promoted_to_fp16 = const()[name = string("const_187_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_5816_cast_fp16 = mul(x = var_5814_cast_fp16_1, y = const_187_promoted_to_fp16)[name = string("op_5816_cast_fp16")];
            int32 var_5818 = const()[name = string("op_5818"), val = int32(-2)];
            bool var_5819_interleave_0 = const()[name = string("op_5819_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_5819_cast_fp16 = concat(axis = var_5818, interleave = var_5819_interleave_0, values = (var_5816_cast_fp16, var_5814_cast_fp16_0))[name = string("op_5819_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_5820_cast_fp16 = mul(x = var_5819_cast_fp16, y = var_866_cast_fp16)[name = string("op_5820_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_93_cast_fp16 = add(x = var_5813_cast_fp16, y = var_5820_cast_fp16)[name = string("query_states_93_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_5826_cast_fp16 = mul(x = var_5802_cast_fp16, y = var_859_cast_fp16)[name = string("op_5826_cast_fp16")];
            tensor<int32, [2]> var_5827_split_sizes_0 = const()[name = string("op_5827_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5827_axis_0 = const()[name = string("op_5827_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_5827_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_5827_cast_fp16_1 = split(axis = var_5827_axis_0, split_sizes = var_5827_split_sizes_0, x = var_5802_cast_fp16)[name = string("op_5827_cast_fp16")];
            fp16 const_188_promoted_to_fp16 = const()[name = string("const_188_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_5829_cast_fp16 = mul(x = var_5827_cast_fp16_1, y = const_188_promoted_to_fp16)[name = string("op_5829_cast_fp16")];
            int32 var_5831 = const()[name = string("op_5831"), val = int32(-2)];
            bool var_5832_interleave_0 = const()[name = string("op_5832_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_5832_cast_fp16 = concat(axis = var_5831, interleave = var_5832_interleave_0, values = (var_5829_cast_fp16, var_5827_cast_fp16_0))[name = string("op_5832_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_5833_cast_fp16 = mul(x = var_5832_cast_fp16, y = var_866_cast_fp16)[name = string("op_5833_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_155_cast_fp16 = add(x = var_5826_cast_fp16, y = var_5833_cast_fp16)[name = string("key_states_155_cast_fp16")];
            tensor<int32, [4]> key_states_157_perm_0 = const()[name = string("key_states_157_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_181 = const()[name = string("expand_dims_181"), val = tensor<int32, [1]>([15])];
            tensor<int32, [1]> expand_dims_182 = const()[name = string("expand_dims_182"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_184 = const()[name = string("expand_dims_184"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_185 = const()[name = string("expand_dims_185"), val = tensor<int32, [1]>([16])];
            int32 concat_123_axis_0 = const()[name = string("concat_123_axis_0"), val = int32(0)];
            bool concat_123_interleave_0 = const()[name = string("concat_123_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_123 = concat(axis = concat_123_axis_0, interleave = concat_123_interleave_0, values = (expand_dims_181, expand_dims_182, position_id, expand_dims_184))[name = string("concat_123")];
            tensor<int32, [1]> concat_124_values1_0 = const()[name = string("concat_124_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_124_values3_0 = const()[name = string("concat_124_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_124_axis_0 = const()[name = string("concat_124_axis_0"), val = int32(0)];
            bool concat_124_interleave_0 = const()[name = string("concat_124_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_124 = concat(axis = concat_124_axis_0, interleave = concat_124_interleave_0, values = (expand_dims_185, concat_124_values1_0, cache_position_end, concat_124_values3_0))[name = string("concat_124")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_16_stride_0 = const()[name = string("key_cache_internal_tensor_assign_16_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_16_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_16_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_16_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_16_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_16_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_16_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_157_cast_fp16 = transpose(perm = key_states_157_perm_0, x = key_states_155_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_16_cast_fp16 = slice_update(begin = concat_123, begin_mask = key_cache_internal_tensor_assign_16_begin_mask_0, end = concat_124, end_mask = key_cache_internal_tensor_assign_16_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_16_squeeze_mask_0, stride = key_cache_internal_tensor_assign_16_stride_0, update = key_states_157_cast_fp16, x = coreml_update_state_84)[name = string("key_cache_internal_tensor_assign_16_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_16_cast_fp16, input = key_cache)[name = string("coreml_update_state_86_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_86 = read_state(input = key_cache)[name = string("coreml_update_state_86")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_16_stride_0 = const()[name = string("value_cache_internal_tensor_assign_16_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_16_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_16_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_16_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_16_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_16_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_16_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_93_cast_fp16 = transpose(perm = value_states_93_perm_0, x = var_5809_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_16_cast_fp16 = slice_update(begin = concat_123, begin_mask = value_cache_internal_tensor_assign_16_begin_mask_0, end = concat_124, end_mask = value_cache_internal_tensor_assign_16_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_16_squeeze_mask_0, stride = value_cache_internal_tensor_assign_16_stride_0, update = value_states_93_cast_fp16, x = coreml_update_state_85)[name = string("value_cache_internal_tensor_assign_16_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_16_cast_fp16, input = value_cache)[name = string("coreml_update_state_87_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_87 = read_state(input = value_cache)[name = string("coreml_update_state_87")];
            tensor<int32, [4]> var_5903_begin_0 = const()[name = string("op_5903_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_5903_end_0 = const()[name = string("op_5903_end_0"), val = tensor<int32, [4]>([16, 2, 2048, 128])];
            tensor<bool, [4]> var_5903_end_mask_0 = const()[name = string("op_5903_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_5903_cast_fp16 = slice_by_index(begin = var_5903_begin_0, end = var_5903_end_0, end_mask = var_5903_end_mask_0, x = coreml_update_state_86)[name = string("op_5903_cast_fp16")];
            tensor<int32, [2]> tile_30 = const()[name = string("tile_30"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5906_axis_0 = const()[name = string("op_5906_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_5906_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_5906_cast_fp16_1 = split(axis = var_5906_axis_0, split_sizes = tile_30, x = var_5903_cast_fp16)[name = string("op_5906_cast_fp16")];
            tensor<int32, [4]> var_5913_begin_0 = const()[name = string("op_5913_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_5913_end_0 = const()[name = string("op_5913_end_0"), val = tensor<int32, [4]>([16, 2, 2048, 128])];
            tensor<bool, [4]> var_5913_end_mask_0 = const()[name = string("op_5913_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_5913_cast_fp16 = slice_by_index(begin = var_5913_begin_0, end = var_5913_end_0, end_mask = var_5913_end_mask_0, x = coreml_update_state_87)[name = string("op_5913_cast_fp16")];
            tensor<int32, [2]> tile_31 = const()[name = string("tile_31"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5916_axis_0 = const()[name = string("op_5916_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_5916_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_5916_cast_fp16_1 = split(axis = var_5916_axis_0, split_sizes = tile_31, x = var_5913_cast_fp16)[name = string("op_5916_cast_fp16")];
            tensor<int32, [2]> var_5919_split_sizes_0 = const()[name = string("op_5919_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_5919_axis_0 = const()[name = string("op_5919_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_5919_0, tensor<fp16, [1, 8, 128, 8]> var_5919_1 = split(axis = var_5919_axis_0, split_sizes = var_5919_split_sizes_0, x = query_states_93_cast_fp16)[name = string("op_5919")];
            bool attn_weights_241_transpose_x_0 = const()[name = string("attn_weights_241_transpose_x_0"), val = bool(false)];
            bool attn_weights_241_transpose_y_0 = const()[name = string("attn_weights_241_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_241_cast_fp16 = matmul(transpose_x = attn_weights_241_transpose_x_0, transpose_y = attn_weights_241_transpose_y_0, x = var_5906_cast_fp16_0, y = var_5919_0)[name = string("attn_weights_241_cast_fp16")];
            fp16 var_5922_to_fp16 = const()[name = string("op_5922_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_243_cast_fp16 = mul(x = attn_weights_241_cast_fp16, y = var_5922_to_fp16)[name = string("attn_weights_243_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_245_cast_fp16 = add(x = attn_weights_243_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_245_cast_fp16")];
            int32 var_5926 = const()[name = string("op_5926"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_247_cast_fp16 = softmax(axis = var_5926, x = attn_weights_245_cast_fp16)[name = string("attn_weights_247_cast_fp16")];
            bool var_5932_transpose_x_1 = const()[name = string("op_5932_transpose_x_1"), val = bool(true)];
            bool var_5932_transpose_y_1 = const()[name = string("op_5932_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_5932_cast_fp16 = matmul(transpose_x = var_5932_transpose_x_1, transpose_y = var_5932_transpose_y_1, x = attn_weights_247_cast_fp16, y = var_5916_cast_fp16_0)[name = string("op_5932_cast_fp16")];
            bool attn_weights_249_transpose_x_0 = const()[name = string("attn_weights_249_transpose_x_0"), val = bool(false)];
            bool attn_weights_249_transpose_y_0 = const()[name = string("attn_weights_249_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_249_cast_fp16 = matmul(transpose_x = attn_weights_249_transpose_x_0, transpose_y = attn_weights_249_transpose_y_0, x = var_5906_cast_fp16_1, y = var_5919_1)[name = string("attn_weights_249_cast_fp16")];
            fp16 var_5934_to_fp16 = const()[name = string("op_5934_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_251_cast_fp16 = mul(x = attn_weights_249_cast_fp16, y = var_5934_to_fp16)[name = string("attn_weights_251_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_253_cast_fp16 = add(x = attn_weights_251_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_253_cast_fp16")];
            int32 var_5938 = const()[name = string("op_5938"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_255_cast_fp16 = softmax(axis = var_5938, x = attn_weights_253_cast_fp16)[name = string("attn_weights_255_cast_fp16")];
            bool attn_output_121_transpose_x_1 = const()[name = string("attn_output_121_transpose_x_1"), val = bool(true)];
            bool attn_output_121_transpose_y_1 = const()[name = string("attn_output_121_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_121_cast_fp16 = matmul(transpose_x = attn_output_121_transpose_x_1, transpose_y = attn_output_121_transpose_y_1, x = attn_weights_255_cast_fp16, y = var_5916_cast_fp16_1)[name = string("attn_output_121_cast_fp16")];
            int32 var_5946 = const()[name = string("op_5946"), val = int32(1)];
            bool attn_output_123_interleave_0 = const()[name = string("attn_output_123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_123_cast_fp16 = concat(axis = var_5946, interleave = attn_output_123_interleave_0, values = (var_5932_cast_fp16, attn_output_121_cast_fp16))[name = string("attn_output_123_cast_fp16")];
            tensor<int32, [4]> var_5950_perm_0 = const()[name = string("op_5950_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_5955 = const()[name = string("op_5955"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_5950_cast_fp16 = transpose(perm = var_5950_perm_0, x = attn_output_123_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_127_cast_fp16 = reshape(shape = var_5955, x = var_5950_cast_fp16)[name = string("attn_output_127_cast_fp16")];
            string hidden_states_153_pad_type_0 = const()[name = string("hidden_states_153_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_153_strides_0 = const()[name = string("hidden_states_153_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_153_pad_0 = const()[name = string("hidden_states_153_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_153_dilations_0 = const()[name = string("hidden_states_153_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_153_groups_0 = const()[name = string("hidden_states_153_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_153_cast_fp16 = conv(dilations = hidden_states_153_dilations_0, groups = hidden_states_153_groups_0, pad = hidden_states_153_pad_0, pad_type = hidden_states_153_pad_type_0, strides = hidden_states_153_strides_0, weight = layers_15_self_attn_o_proj_weight_cast_fp16, x = attn_output_127_cast_fp16)[name = string("hidden_states_153_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_155_cast_fp16 = add(x = hidden_states_149_cast_fp16, y = hidden_states_153_cast_fp16)[name = string("hidden_states_155_cast_fp16")];
            int32 var_5981 = const()[name = string("op_5981"), val = int32(1)];
            fp16 const_193_promoted_to_fp16 = const()[name = string("const_193_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_5983_cast_fp16 = mul(x = hidden_states_155_cast_fp16, y = const_193_promoted_to_fp16)[name = string("op_5983_cast_fp16")];
            bool doubled_125_interleave_0 = const()[name = string("doubled_125_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_125_cast_fp16 = concat(axis = var_5981, interleave = doubled_125_interleave_0, values = (hidden_states_155_cast_fp16, var_5983_cast_fp16))[name = string("doubled_125_cast_fp16")];
            tensor<int32, [1]> out_63_axes_0 = const()[name = string("out_63_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_63_gamma_0_to_fp16 = const()[name = string("out_63_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1433299392)))];
            fp16 var_5993_to_fp16 = const()[name = string("op_5993_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_5993_to_fp16, gamma = out_63_gamma_0_to_fp16, x = doubled_125_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<int32, [2]> var_6004_split_sizes_0 = const()[name = string("op_6004_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_6004_axis_0 = const()[name = string("op_6004_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_6004_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_6004_cast_fp16_1 = split(axis = var_6004_axis_0, split_sizes = var_6004_split_sizes_0, x = out_63_cast_fp16)[name = string("op_6004_cast_fp16")];
            string input_31_pad_type_0 = const()[name = string("input_31_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_31_strides_0 = const()[name = string("input_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_31_pad_0 = const()[name = string("input_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_31_dilations_0 = const()[name = string("input_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_31_groups_0 = const()[name = string("input_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_31_cast_fp16 = conv(dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = layers_15_mlp_gate_proj_weight_cast_fp16, x = var_6004_cast_fp16_0)[name = string("input_31_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_6021_cast_fp16 = silu(x = input_31_cast_fp16)[name = string("op_6021_cast_fp16")];
            string var_6027_pad_type_0 = const()[name = string("op_6027_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6027_strides_0 = const()[name = string("op_6027_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6027_pad_0 = const()[name = string("op_6027_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6027_dilations_0 = const()[name = string("op_6027_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6027_groups_0 = const()[name = string("op_6027_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_6027_cast_fp16 = conv(dilations = var_6027_dilations_0, groups = var_6027_groups_0, pad = var_6027_pad_0, pad_type = var_6027_pad_type_0, strides = var_6027_strides_0, weight = layers_15_mlp_up_proj_weight_cast_fp16, x = var_6004_cast_fp16_0)[name = string("op_6027_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_159_cast_fp16 = mul(x = var_6021_cast_fp16, y = var_6027_cast_fp16)[name = string("x_159_cast_fp16")];
            string hidden_states_157_pad_type_0 = const()[name = string("hidden_states_157_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_157_strides_0 = const()[name = string("hidden_states_157_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_157_pad_0 = const()[name = string("hidden_states_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_157_dilations_0 = const()[name = string("hidden_states_157_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_157_groups_0 = const()[name = string("hidden_states_157_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_157_cast_fp16 = conv(dilations = hidden_states_157_dilations_0, groups = hidden_states_157_groups_0, pad = hidden_states_157_pad_0, pad_type = hidden_states_157_pad_type_0, strides = hidden_states_157_strides_0, weight = layers_15_mlp_down_proj_weight_cast_fp16, x = x_159_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_159_cast_fp16 = add(x = hidden_states_155_cast_fp16, y = hidden_states_157_cast_fp16)[name = string("hidden_states_159_cast_fp16")];
            int32 var_6043 = const()[name = string("op_6043"), val = int32(1)];
            fp16 const_195_promoted_to_fp16 = const()[name = string("const_195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_6045_cast_fp16 = mul(x = hidden_states_159_cast_fp16, y = const_195_promoted_to_fp16)[name = string("op_6045_cast_fp16")];
            bool doubled_129_interleave_0 = const()[name = string("doubled_129_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_129_cast_fp16 = concat(axis = var_6043, interleave = doubled_129_interleave_0, values = (hidden_states_159_cast_fp16, var_6045_cast_fp16))[name = string("doubled_129_cast_fp16")];
            tensor<int32, [1]> out_65_axes_0 = const()[name = string("out_65_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_65_gamma_0_to_fp16 = const()[name = string("out_65_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1433307648)))];
            fp16 var_6055_to_fp16 = const()[name = string("op_6055_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_6055_to_fp16, gamma = out_65_gamma_0_to_fp16, x = doubled_129_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<int32, [2]> var_6066_split_sizes_0 = const()[name = string("op_6066_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_6066_axis_0 = const()[name = string("op_6066_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_6066_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_6066_cast_fp16_1 = split(axis = var_6066_axis_0, split_sizes = var_6066_split_sizes_0, x = out_65_cast_fp16)[name = string("op_6066_cast_fp16")];
            string query_states_97_pad_type_0 = const()[name = string("query_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_97_strides_0 = const()[name = string("query_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_97_pad_0 = const()[name = string("query_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_97_dilations_0 = const()[name = string("query_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_97_groups_0 = const()[name = string("query_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_97_cast_fp16 = conv(dilations = query_states_97_dilations_0, groups = query_states_97_groups_0, pad = query_states_97_pad_0, pad_type = query_states_97_pad_type_0, strides = query_states_97_strides_0, weight = layers_16_self_attn_q_proj_weight_cast_fp16, x = var_6066_cast_fp16_0)[name = string("query_states_97_cast_fp16")];
            string key_states_161_pad_type_0 = const()[name = string("key_states_161_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_161_strides_0 = const()[name = string("key_states_161_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_161_pad_0 = const()[name = string("key_states_161_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_161_dilations_0 = const()[name = string("key_states_161_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_161_groups_0 = const()[name = string("key_states_161_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_161_cast_fp16 = conv(dilations = key_states_161_dilations_0, groups = key_states_161_groups_0, pad = key_states_161_pad_0, pad_type = key_states_161_pad_type_0, strides = key_states_161_strides_0, weight = layers_16_self_attn_k_proj_weight_cast_fp16, x = var_6066_cast_fp16_0)[name = string("key_states_161_cast_fp16")];
            string value_states_97_pad_type_0 = const()[name = string("value_states_97_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_97_strides_0 = const()[name = string("value_states_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_97_pad_0 = const()[name = string("value_states_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_97_dilations_0 = const()[name = string("value_states_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_97_groups_0 = const()[name = string("value_states_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_97_cast_fp16 = conv(dilations = value_states_97_dilations_0, groups = value_states_97_groups_0, pad = value_states_97_pad_0, pad_type = value_states_97_pad_type_0, strides = value_states_97_strides_0, weight = layers_16_self_attn_v_proj_weight_cast_fp16, x = var_6066_cast_fp16_0)[name = string("value_states_97_cast_fp16")];
            tensor<int32, [4]> var_6118 = const()[name = string("op_6118"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_161_cast_fp16 = reshape(shape = var_6118, x = query_states_97_cast_fp16)[name = string("x_161_cast_fp16")];
            tensor<int32, [4]> var_6122 = const()[name = string("op_6122"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_6123_cast_fp16 = reshape(shape = var_6122, x = key_states_161_cast_fp16)[name = string("op_6123_cast_fp16")];
            tensor<int32, [4]> var_6129 = const()[name = string("op_6129"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_6130_cast_fp16 = reshape(shape = var_6129, x = value_states_97_cast_fp16)[name = string("op_6130_cast_fp16")];
            tensor<int32, [4]> value_states_99_perm_0 = const()[name = string("value_states_99_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_6134_cast_fp16 = mul(x = x_161_cast_fp16, y = var_859_cast_fp16)[name = string("op_6134_cast_fp16")];
            tensor<int32, [2]> var_6135_split_sizes_0 = const()[name = string("op_6135_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_6135_axis_0 = const()[name = string("op_6135_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_6135_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_6135_cast_fp16_1 = split(axis = var_6135_axis_0, split_sizes = var_6135_split_sizes_0, x = x_161_cast_fp16)[name = string("op_6135_cast_fp16")];
            fp16 const_199_promoted_to_fp16 = const()[name = string("const_199_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_6137_cast_fp16 = mul(x = var_6135_cast_fp16_1, y = const_199_promoted_to_fp16)[name = string("op_6137_cast_fp16")];
            int32 var_6139 = const()[name = string("op_6139"), val = int32(-2)];
            bool var_6140_interleave_0 = const()[name = string("op_6140_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_6140_cast_fp16 = concat(axis = var_6139, interleave = var_6140_interleave_0, values = (var_6137_cast_fp16, var_6135_cast_fp16_0))[name = string("op_6140_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_6141_cast_fp16 = mul(x = var_6140_cast_fp16, y = var_866_cast_fp16)[name = string("op_6141_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_99_cast_fp16 = add(x = var_6134_cast_fp16, y = var_6141_cast_fp16)[name = string("query_states_99_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_6147_cast_fp16 = mul(x = var_6123_cast_fp16, y = var_859_cast_fp16)[name = string("op_6147_cast_fp16")];
            tensor<int32, [2]> var_6148_split_sizes_0 = const()[name = string("op_6148_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_6148_axis_0 = const()[name = string("op_6148_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_6148_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_6148_cast_fp16_1 = split(axis = var_6148_axis_0, split_sizes = var_6148_split_sizes_0, x = var_6123_cast_fp16)[name = string("op_6148_cast_fp16")];
            fp16 const_200_promoted_to_fp16 = const()[name = string("const_200_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_6150_cast_fp16 = mul(x = var_6148_cast_fp16_1, y = const_200_promoted_to_fp16)[name = string("op_6150_cast_fp16")];
            int32 var_6152 = const()[name = string("op_6152"), val = int32(-2)];
            bool var_6153_interleave_0 = const()[name = string("op_6153_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_6153_cast_fp16 = concat(axis = var_6152, interleave = var_6153_interleave_0, values = (var_6150_cast_fp16, var_6148_cast_fp16_0))[name = string("op_6153_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_6154_cast_fp16 = mul(x = var_6153_cast_fp16, y = var_866_cast_fp16)[name = string("op_6154_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_165_cast_fp16 = add(x = var_6147_cast_fp16, y = var_6154_cast_fp16)[name = string("key_states_165_cast_fp16")];
            tensor<int32, [4]> key_states_167_perm_0 = const()[name = string("key_states_167_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_193 = const()[name = string("expand_dims_193"), val = tensor<int32, [1]>([16])];
            tensor<int32, [1]> expand_dims_194 = const()[name = string("expand_dims_194"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_196 = const()[name = string("expand_dims_196"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_197 = const()[name = string("expand_dims_197"), val = tensor<int32, [1]>([17])];
            int32 concat_131_axis_0 = const()[name = string("concat_131_axis_0"), val = int32(0)];
            bool concat_131_interleave_0 = const()[name = string("concat_131_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_131 = concat(axis = concat_131_axis_0, interleave = concat_131_interleave_0, values = (expand_dims_193, expand_dims_194, position_id, expand_dims_196))[name = string("concat_131")];
            tensor<int32, [1]> concat_132_values1_0 = const()[name = string("concat_132_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_132_values3_0 = const()[name = string("concat_132_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_132_axis_0 = const()[name = string("concat_132_axis_0"), val = int32(0)];
            bool concat_132_interleave_0 = const()[name = string("concat_132_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_132 = concat(axis = concat_132_axis_0, interleave = concat_132_interleave_0, values = (expand_dims_197, concat_132_values1_0, cache_position_end, concat_132_values3_0))[name = string("concat_132")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_17_stride_0 = const()[name = string("key_cache_internal_tensor_assign_17_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_17_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_17_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_17_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_17_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_17_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_17_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_167_cast_fp16 = transpose(perm = key_states_167_perm_0, x = key_states_165_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_17_cast_fp16 = slice_update(begin = concat_131, begin_mask = key_cache_internal_tensor_assign_17_begin_mask_0, end = concat_132, end_mask = key_cache_internal_tensor_assign_17_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_17_squeeze_mask_0, stride = key_cache_internal_tensor_assign_17_stride_0, update = key_states_167_cast_fp16, x = coreml_update_state_86)[name = string("key_cache_internal_tensor_assign_17_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_17_cast_fp16, input = key_cache)[name = string("coreml_update_state_88_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_88 = read_state(input = key_cache)[name = string("coreml_update_state_88")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_17_stride_0 = const()[name = string("value_cache_internal_tensor_assign_17_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_17_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_17_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_17_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_17_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_17_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_17_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_99_cast_fp16 = transpose(perm = value_states_99_perm_0, x = var_6130_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_17_cast_fp16 = slice_update(begin = concat_131, begin_mask = value_cache_internal_tensor_assign_17_begin_mask_0, end = concat_132, end_mask = value_cache_internal_tensor_assign_17_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_17_squeeze_mask_0, stride = value_cache_internal_tensor_assign_17_stride_0, update = value_states_99_cast_fp16, x = coreml_update_state_87)[name = string("value_cache_internal_tensor_assign_17_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_17_cast_fp16, input = value_cache)[name = string("coreml_update_state_89_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_89 = read_state(input = value_cache)[name = string("coreml_update_state_89")];
            tensor<int32, [4]> var_6224_begin_0 = const()[name = string("op_6224_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6224_end_0 = const()[name = string("op_6224_end_0"), val = tensor<int32, [4]>([17, 2, 2048, 128])];
            tensor<bool, [4]> var_6224_end_mask_0 = const()[name = string("op_6224_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_6224_cast_fp16 = slice_by_index(begin = var_6224_begin_0, end = var_6224_end_0, end_mask = var_6224_end_mask_0, x = coreml_update_state_88)[name = string("op_6224_cast_fp16")];
            tensor<int32, [2]> tile_32 = const()[name = string("tile_32"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6227_axis_0 = const()[name = string("op_6227_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_6227_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_6227_cast_fp16_1 = split(axis = var_6227_axis_0, split_sizes = tile_32, x = var_6224_cast_fp16)[name = string("op_6227_cast_fp16")];
            tensor<int32, [4]> var_6234_begin_0 = const()[name = string("op_6234_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6234_end_0 = const()[name = string("op_6234_end_0"), val = tensor<int32, [4]>([17, 2, 2048, 128])];
            tensor<bool, [4]> var_6234_end_mask_0 = const()[name = string("op_6234_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_6234_cast_fp16 = slice_by_index(begin = var_6234_begin_0, end = var_6234_end_0, end_mask = var_6234_end_mask_0, x = coreml_update_state_89)[name = string("op_6234_cast_fp16")];
            tensor<int32, [2]> tile_33 = const()[name = string("tile_33"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6237_axis_0 = const()[name = string("op_6237_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_6237_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_6237_cast_fp16_1 = split(axis = var_6237_axis_0, split_sizes = tile_33, x = var_6234_cast_fp16)[name = string("op_6237_cast_fp16")];
            tensor<int32, [2]> var_6240_split_sizes_0 = const()[name = string("op_6240_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_6240_axis_0 = const()[name = string("op_6240_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_6240_0, tensor<fp16, [1, 8, 128, 8]> var_6240_1 = split(axis = var_6240_axis_0, split_sizes = var_6240_split_sizes_0, x = query_states_99_cast_fp16)[name = string("op_6240")];
            bool attn_weights_257_transpose_x_0 = const()[name = string("attn_weights_257_transpose_x_0"), val = bool(false)];
            bool attn_weights_257_transpose_y_0 = const()[name = string("attn_weights_257_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_257_cast_fp16 = matmul(transpose_x = attn_weights_257_transpose_x_0, transpose_y = attn_weights_257_transpose_y_0, x = var_6227_cast_fp16_0, y = var_6240_0)[name = string("attn_weights_257_cast_fp16")];
            fp16 var_6243_to_fp16 = const()[name = string("op_6243_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_259_cast_fp16 = mul(x = attn_weights_257_cast_fp16, y = var_6243_to_fp16)[name = string("attn_weights_259_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_261_cast_fp16 = add(x = attn_weights_259_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_261_cast_fp16")];
            int32 var_6247 = const()[name = string("op_6247"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_263_cast_fp16 = softmax(axis = var_6247, x = attn_weights_261_cast_fp16)[name = string("attn_weights_263_cast_fp16")];
            bool var_6253_transpose_x_1 = const()[name = string("op_6253_transpose_x_1"), val = bool(true)];
            bool var_6253_transpose_y_1 = const()[name = string("op_6253_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_6253_cast_fp16 = matmul(transpose_x = var_6253_transpose_x_1, transpose_y = var_6253_transpose_y_1, x = attn_weights_263_cast_fp16, y = var_6237_cast_fp16_0)[name = string("op_6253_cast_fp16")];
            bool attn_weights_265_transpose_x_0 = const()[name = string("attn_weights_265_transpose_x_0"), val = bool(false)];
            bool attn_weights_265_transpose_y_0 = const()[name = string("attn_weights_265_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_265_cast_fp16 = matmul(transpose_x = attn_weights_265_transpose_x_0, transpose_y = attn_weights_265_transpose_y_0, x = var_6227_cast_fp16_1, y = var_6240_1)[name = string("attn_weights_265_cast_fp16")];
            fp16 var_6255_to_fp16 = const()[name = string("op_6255_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_267_cast_fp16 = mul(x = attn_weights_265_cast_fp16, y = var_6255_to_fp16)[name = string("attn_weights_267_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_269_cast_fp16 = add(x = attn_weights_267_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_269_cast_fp16")];
            int32 var_6259 = const()[name = string("op_6259"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_271_cast_fp16 = softmax(axis = var_6259, x = attn_weights_269_cast_fp16)[name = string("attn_weights_271_cast_fp16")];
            bool attn_output_129_transpose_x_1 = const()[name = string("attn_output_129_transpose_x_1"), val = bool(true)];
            bool attn_output_129_transpose_y_1 = const()[name = string("attn_output_129_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_129_cast_fp16 = matmul(transpose_x = attn_output_129_transpose_x_1, transpose_y = attn_output_129_transpose_y_1, x = attn_weights_271_cast_fp16, y = var_6237_cast_fp16_1)[name = string("attn_output_129_cast_fp16")];
            int32 var_6267 = const()[name = string("op_6267"), val = int32(1)];
            bool attn_output_131_interleave_0 = const()[name = string("attn_output_131_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_131_cast_fp16 = concat(axis = var_6267, interleave = attn_output_131_interleave_0, values = (var_6253_cast_fp16, attn_output_129_cast_fp16))[name = string("attn_output_131_cast_fp16")];
            tensor<int32, [4]> var_6271_perm_0 = const()[name = string("op_6271_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_6276 = const()[name = string("op_6276"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_6271_cast_fp16 = transpose(perm = var_6271_perm_0, x = attn_output_131_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_135_cast_fp16 = reshape(shape = var_6276, x = var_6271_cast_fp16)[name = string("attn_output_135_cast_fp16")];
            string hidden_states_163_pad_type_0 = const()[name = string("hidden_states_163_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_163_strides_0 = const()[name = string("hidden_states_163_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_163_pad_0 = const()[name = string("hidden_states_163_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_163_dilations_0 = const()[name = string("hidden_states_163_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_163_groups_0 = const()[name = string("hidden_states_163_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_163_cast_fp16 = conv(dilations = hidden_states_163_dilations_0, groups = hidden_states_163_groups_0, pad = hidden_states_163_pad_0, pad_type = hidden_states_163_pad_type_0, strides = hidden_states_163_strides_0, weight = layers_16_self_attn_o_proj_weight_cast_fp16, x = attn_output_135_cast_fp16)[name = string("hidden_states_163_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_165_cast_fp16 = add(x = hidden_states_159_cast_fp16, y = hidden_states_163_cast_fp16)[name = string("hidden_states_165_cast_fp16")];
            int32 var_6302 = const()[name = string("op_6302"), val = int32(1)];
            fp16 const_205_promoted_to_fp16 = const()[name = string("const_205_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_6304_cast_fp16 = mul(x = hidden_states_165_cast_fp16, y = const_205_promoted_to_fp16)[name = string("op_6304_cast_fp16")];
            bool doubled_133_interleave_0 = const()[name = string("doubled_133_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_133_cast_fp16 = concat(axis = var_6302, interleave = doubled_133_interleave_0, values = (hidden_states_165_cast_fp16, var_6304_cast_fp16))[name = string("doubled_133_cast_fp16")];
            tensor<int32, [1]> out_67_axes_0 = const()[name = string("out_67_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_67_gamma_0_to_fp16 = const()[name = string("out_67_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1433315904)))];
            fp16 var_6314_to_fp16 = const()[name = string("op_6314_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_6314_to_fp16, gamma = out_67_gamma_0_to_fp16, x = doubled_133_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<int32, [2]> var_6325_split_sizes_0 = const()[name = string("op_6325_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_6325_axis_0 = const()[name = string("op_6325_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_6325_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_6325_cast_fp16_1 = split(axis = var_6325_axis_0, split_sizes = var_6325_split_sizes_0, x = out_67_cast_fp16)[name = string("op_6325_cast_fp16")];
            string input_33_pad_type_0 = const()[name = string("input_33_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_33_strides_0 = const()[name = string("input_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_33_pad_0 = const()[name = string("input_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_33_dilations_0 = const()[name = string("input_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_33_groups_0 = const()[name = string("input_33_groups_0"), val = int32(1)];
            tensor<fp16, [6144, 2048, 1, 1]> layers_16_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_16_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1433324160)))];
            tensor<fp16, [1, 6144, 1, 8]> input_33_cast_fp16 = conv(dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = layers_16_mlp_gate_proj_weight_to_fp16, x = var_6325_cast_fp16_0)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_6342_cast_fp16 = silu(x = input_33_cast_fp16)[name = string("op_6342_cast_fp16")];
            string var_6348_pad_type_0 = const()[name = string("op_6348_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6348_strides_0 = const()[name = string("op_6348_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6348_pad_0 = const()[name = string("op_6348_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6348_dilations_0 = const()[name = string("op_6348_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6348_groups_0 = const()[name = string("op_6348_groups_0"), val = int32(1)];
            tensor<fp16, [6144, 2048, 1, 1]> layers_16_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_16_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1458490048)))];
            tensor<fp16, [1, 6144, 1, 8]> var_6348_cast_fp16 = conv(dilations = var_6348_dilations_0, groups = var_6348_groups_0, pad = var_6348_pad_0, pad_type = var_6348_pad_type_0, strides = var_6348_strides_0, weight = layers_16_mlp_up_proj_weight_to_fp16, x = var_6325_cast_fp16_0)[name = string("op_6348_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_169_cast_fp16 = mul(x = var_6342_cast_fp16, y = var_6348_cast_fp16)[name = string("x_169_cast_fp16")];
            string hidden_states_167_pad_type_0 = const()[name = string("hidden_states_167_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_167_strides_0 = const()[name = string("hidden_states_167_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_167_pad_0 = const()[name = string("hidden_states_167_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_167_dilations_0 = const()[name = string("hidden_states_167_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_167_groups_0 = const()[name = string("hidden_states_167_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_167_cast_fp16 = conv(dilations = hidden_states_167_dilations_0, groups = hidden_states_167_groups_0, pad = hidden_states_167_pad_0, pad_type = hidden_states_167_pad_type_0, strides = hidden_states_167_strides_0, weight = layers_16_mlp_down_proj_weight_cast_fp16, x = x_169_cast_fp16)[name = string("hidden_states_167_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_169_cast_fp16 = add(x = hidden_states_165_cast_fp16, y = hidden_states_167_cast_fp16)[name = string("hidden_states_169_cast_fp16")];
            int32 var_6364 = const()[name = string("op_6364"), val = int32(1)];
            fp16 const_207_promoted_to_fp16 = const()[name = string("const_207_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_6366_cast_fp16 = mul(x = hidden_states_169_cast_fp16, y = const_207_promoted_to_fp16)[name = string("op_6366_cast_fp16")];
            bool doubled_137_interleave_0 = const()[name = string("doubled_137_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_137_cast_fp16 = concat(axis = var_6364, interleave = doubled_137_interleave_0, values = (hidden_states_169_cast_fp16, var_6366_cast_fp16))[name = string("doubled_137_cast_fp16")];
            tensor<int32, [1]> out_69_axes_0 = const()[name = string("out_69_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_69_gamma_0_to_fp16 = const()[name = string("out_69_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1483655936)))];
            fp16 var_6376_to_fp16 = const()[name = string("op_6376_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_6376_to_fp16, gamma = out_69_gamma_0_to_fp16, x = doubled_137_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<int32, [2]> var_6387_split_sizes_0 = const()[name = string("op_6387_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_6387_axis_0 = const()[name = string("op_6387_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_6387_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_6387_cast_fp16_1 = split(axis = var_6387_axis_0, split_sizes = var_6387_split_sizes_0, x = out_69_cast_fp16)[name = string("op_6387_cast_fp16")];
            string query_states_103_pad_type_0 = const()[name = string("query_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_103_strides_0 = const()[name = string("query_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_103_pad_0 = const()[name = string("query_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_103_dilations_0 = const()[name = string("query_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_103_groups_0 = const()[name = string("query_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_103_cast_fp16 = conv(dilations = query_states_103_dilations_0, groups = query_states_103_groups_0, pad = query_states_103_pad_0, pad_type = query_states_103_pad_type_0, strides = query_states_103_strides_0, weight = layers_17_self_attn_q_proj_weight_cast_fp16, x = var_6387_cast_fp16_0)[name = string("query_states_103_cast_fp16")];
            string key_states_171_pad_type_0 = const()[name = string("key_states_171_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_171_strides_0 = const()[name = string("key_states_171_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_171_pad_0 = const()[name = string("key_states_171_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_171_dilations_0 = const()[name = string("key_states_171_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_171_groups_0 = const()[name = string("key_states_171_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_171_cast_fp16 = conv(dilations = key_states_171_dilations_0, groups = key_states_171_groups_0, pad = key_states_171_pad_0, pad_type = key_states_171_pad_type_0, strides = key_states_171_strides_0, weight = layers_17_self_attn_k_proj_weight_cast_fp16, x = var_6387_cast_fp16_0)[name = string("key_states_171_cast_fp16")];
            string value_states_103_pad_type_0 = const()[name = string("value_states_103_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_103_strides_0 = const()[name = string("value_states_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_103_pad_0 = const()[name = string("value_states_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_103_dilations_0 = const()[name = string("value_states_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_103_groups_0 = const()[name = string("value_states_103_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_103_cast_fp16 = conv(dilations = value_states_103_dilations_0, groups = value_states_103_groups_0, pad = value_states_103_pad_0, pad_type = value_states_103_pad_type_0, strides = value_states_103_strides_0, weight = layers_17_self_attn_v_proj_weight_cast_fp16, x = var_6387_cast_fp16_0)[name = string("value_states_103_cast_fp16")];
            tensor<int32, [4]> var_6439 = const()[name = string("op_6439"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_171_cast_fp16 = reshape(shape = var_6439, x = query_states_103_cast_fp16)[name = string("x_171_cast_fp16")];
            tensor<int32, [4]> var_6443 = const()[name = string("op_6443"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_6444_cast_fp16 = reshape(shape = var_6443, x = key_states_171_cast_fp16)[name = string("op_6444_cast_fp16")];
            tensor<int32, [4]> var_6450 = const()[name = string("op_6450"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_6451_cast_fp16 = reshape(shape = var_6450, x = value_states_103_cast_fp16)[name = string("op_6451_cast_fp16")];
            tensor<int32, [4]> value_states_105_perm_0 = const()[name = string("value_states_105_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_6455_cast_fp16 = mul(x = x_171_cast_fp16, y = var_859_cast_fp16)[name = string("op_6455_cast_fp16")];
            tensor<int32, [2]> var_6456_split_sizes_0 = const()[name = string("op_6456_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_6456_axis_0 = const()[name = string("op_6456_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_6456_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_6456_cast_fp16_1 = split(axis = var_6456_axis_0, split_sizes = var_6456_split_sizes_0, x = x_171_cast_fp16)[name = string("op_6456_cast_fp16")];
            fp16 const_211_promoted_to_fp16 = const()[name = string("const_211_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_6458_cast_fp16 = mul(x = var_6456_cast_fp16_1, y = const_211_promoted_to_fp16)[name = string("op_6458_cast_fp16")];
            int32 var_6460 = const()[name = string("op_6460"), val = int32(-2)];
            bool var_6461_interleave_0 = const()[name = string("op_6461_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_6461_cast_fp16 = concat(axis = var_6460, interleave = var_6461_interleave_0, values = (var_6458_cast_fp16, var_6456_cast_fp16_0))[name = string("op_6461_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_6462_cast_fp16 = mul(x = var_6461_cast_fp16, y = var_866_cast_fp16)[name = string("op_6462_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_105_cast_fp16 = add(x = var_6455_cast_fp16, y = var_6462_cast_fp16)[name = string("query_states_105_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_6468_cast_fp16 = mul(x = var_6444_cast_fp16, y = var_859_cast_fp16)[name = string("op_6468_cast_fp16")];
            tensor<int32, [2]> var_6469_split_sizes_0 = const()[name = string("op_6469_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_6469_axis_0 = const()[name = string("op_6469_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_6469_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_6469_cast_fp16_1 = split(axis = var_6469_axis_0, split_sizes = var_6469_split_sizes_0, x = var_6444_cast_fp16)[name = string("op_6469_cast_fp16")];
            fp16 const_212_promoted_to_fp16 = const()[name = string("const_212_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_6471_cast_fp16 = mul(x = var_6469_cast_fp16_1, y = const_212_promoted_to_fp16)[name = string("op_6471_cast_fp16")];
            int32 var_6473 = const()[name = string("op_6473"), val = int32(-2)];
            bool var_6474_interleave_0 = const()[name = string("op_6474_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_6474_cast_fp16 = concat(axis = var_6473, interleave = var_6474_interleave_0, values = (var_6471_cast_fp16, var_6469_cast_fp16_0))[name = string("op_6474_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_6475_cast_fp16 = mul(x = var_6474_cast_fp16, y = var_866_cast_fp16)[name = string("op_6475_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_175_cast_fp16 = add(x = var_6468_cast_fp16, y = var_6475_cast_fp16)[name = string("key_states_175_cast_fp16")];
            tensor<int32, [4]> key_states_177_perm_0 = const()[name = string("key_states_177_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_205 = const()[name = string("expand_dims_205"), val = tensor<int32, [1]>([17])];
            tensor<int32, [1]> expand_dims_206 = const()[name = string("expand_dims_206"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_208 = const()[name = string("expand_dims_208"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_209 = const()[name = string("expand_dims_209"), val = tensor<int32, [1]>([18])];
            int32 concat_139_axis_0 = const()[name = string("concat_139_axis_0"), val = int32(0)];
            bool concat_139_interleave_0 = const()[name = string("concat_139_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_139 = concat(axis = concat_139_axis_0, interleave = concat_139_interleave_0, values = (expand_dims_205, expand_dims_206, position_id, expand_dims_208))[name = string("concat_139")];
            tensor<int32, [1]> concat_140_values1_0 = const()[name = string("concat_140_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_140_values3_0 = const()[name = string("concat_140_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_140_axis_0 = const()[name = string("concat_140_axis_0"), val = int32(0)];
            bool concat_140_interleave_0 = const()[name = string("concat_140_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_140 = concat(axis = concat_140_axis_0, interleave = concat_140_interleave_0, values = (expand_dims_209, concat_140_values1_0, cache_position_end, concat_140_values3_0))[name = string("concat_140")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_18_stride_0 = const()[name = string("key_cache_internal_tensor_assign_18_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_18_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_18_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_18_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_18_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_18_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_18_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_177_cast_fp16 = transpose(perm = key_states_177_perm_0, x = key_states_175_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_18_cast_fp16 = slice_update(begin = concat_139, begin_mask = key_cache_internal_tensor_assign_18_begin_mask_0, end = concat_140, end_mask = key_cache_internal_tensor_assign_18_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_18_squeeze_mask_0, stride = key_cache_internal_tensor_assign_18_stride_0, update = key_states_177_cast_fp16, x = coreml_update_state_88)[name = string("key_cache_internal_tensor_assign_18_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_18_cast_fp16, input = key_cache)[name = string("coreml_update_state_90_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_90 = read_state(input = key_cache)[name = string("coreml_update_state_90")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_18_stride_0 = const()[name = string("value_cache_internal_tensor_assign_18_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_18_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_18_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_18_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_18_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_18_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_18_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_105_cast_fp16 = transpose(perm = value_states_105_perm_0, x = var_6451_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_18_cast_fp16 = slice_update(begin = concat_139, begin_mask = value_cache_internal_tensor_assign_18_begin_mask_0, end = concat_140, end_mask = value_cache_internal_tensor_assign_18_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_18_squeeze_mask_0, stride = value_cache_internal_tensor_assign_18_stride_0, update = value_states_105_cast_fp16, x = coreml_update_state_89)[name = string("value_cache_internal_tensor_assign_18_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_18_cast_fp16, input = value_cache)[name = string("coreml_update_state_91_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_91 = read_state(input = value_cache)[name = string("coreml_update_state_91")];
            tensor<int32, [4]> var_6545_begin_0 = const()[name = string("op_6545_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_6545_end_0 = const()[name = string("op_6545_end_0"), val = tensor<int32, [4]>([18, 2, 2048, 128])];
            tensor<bool, [4]> var_6545_end_mask_0 = const()[name = string("op_6545_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_6545_cast_fp16 = slice_by_index(begin = var_6545_begin_0, end = var_6545_end_0, end_mask = var_6545_end_mask_0, x = coreml_update_state_90)[name = string("op_6545_cast_fp16")];
            tensor<int32, [2]> tile_34 = const()[name = string("tile_34"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6548_axis_0 = const()[name = string("op_6548_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_6548_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_6548_cast_fp16_1 = split(axis = var_6548_axis_0, split_sizes = tile_34, x = var_6545_cast_fp16)[name = string("op_6548_cast_fp16")];
            tensor<int32, [4]> var_6555_begin_0 = const()[name = string("op_6555_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_6555_end_0 = const()[name = string("op_6555_end_0"), val = tensor<int32, [4]>([18, 2, 2048, 128])];
            tensor<bool, [4]> var_6555_end_mask_0 = const()[name = string("op_6555_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_6555_cast_fp16 = slice_by_index(begin = var_6555_begin_0, end = var_6555_end_0, end_mask = var_6555_end_mask_0, x = coreml_update_state_91)[name = string("op_6555_cast_fp16")];
            tensor<int32, [2]> tile_35 = const()[name = string("tile_35"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6558_axis_0 = const()[name = string("op_6558_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_6558_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_6558_cast_fp16_1 = split(axis = var_6558_axis_0, split_sizes = tile_35, x = var_6555_cast_fp16)[name = string("op_6558_cast_fp16")];
            tensor<int32, [2]> var_6561_split_sizes_0 = const()[name = string("op_6561_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_6561_axis_0 = const()[name = string("op_6561_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_6561_0, tensor<fp16, [1, 8, 128, 8]> var_6561_1 = split(axis = var_6561_axis_0, split_sizes = var_6561_split_sizes_0, x = query_states_105_cast_fp16)[name = string("op_6561")];
            bool attn_weights_273_transpose_x_0 = const()[name = string("attn_weights_273_transpose_x_0"), val = bool(false)];
            bool attn_weights_273_transpose_y_0 = const()[name = string("attn_weights_273_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_273_cast_fp16 = matmul(transpose_x = attn_weights_273_transpose_x_0, transpose_y = attn_weights_273_transpose_y_0, x = var_6548_cast_fp16_0, y = var_6561_0)[name = string("attn_weights_273_cast_fp16")];
            fp16 var_6564_to_fp16 = const()[name = string("op_6564_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_275_cast_fp16 = mul(x = attn_weights_273_cast_fp16, y = var_6564_to_fp16)[name = string("attn_weights_275_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_277_cast_fp16 = add(x = attn_weights_275_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_277_cast_fp16")];
            int32 var_6568 = const()[name = string("op_6568"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_279_cast_fp16 = softmax(axis = var_6568, x = attn_weights_277_cast_fp16)[name = string("attn_weights_279_cast_fp16")];
            bool var_6574_transpose_x_1 = const()[name = string("op_6574_transpose_x_1"), val = bool(true)];
            bool var_6574_transpose_y_1 = const()[name = string("op_6574_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_6574_cast_fp16 = matmul(transpose_x = var_6574_transpose_x_1, transpose_y = var_6574_transpose_y_1, x = attn_weights_279_cast_fp16, y = var_6558_cast_fp16_0)[name = string("op_6574_cast_fp16")];
            bool attn_weights_281_transpose_x_0 = const()[name = string("attn_weights_281_transpose_x_0"), val = bool(false)];
            bool attn_weights_281_transpose_y_0 = const()[name = string("attn_weights_281_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_281_cast_fp16 = matmul(transpose_x = attn_weights_281_transpose_x_0, transpose_y = attn_weights_281_transpose_y_0, x = var_6548_cast_fp16_1, y = var_6561_1)[name = string("attn_weights_281_cast_fp16")];
            fp16 var_6576_to_fp16 = const()[name = string("op_6576_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_283_cast_fp16 = mul(x = attn_weights_281_cast_fp16, y = var_6576_to_fp16)[name = string("attn_weights_283_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_285_cast_fp16 = add(x = attn_weights_283_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_285_cast_fp16")];
            int32 var_6580 = const()[name = string("op_6580"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_287_cast_fp16 = softmax(axis = var_6580, x = attn_weights_285_cast_fp16)[name = string("attn_weights_287_cast_fp16")];
            bool attn_output_137_transpose_x_1 = const()[name = string("attn_output_137_transpose_x_1"), val = bool(true)];
            bool attn_output_137_transpose_y_1 = const()[name = string("attn_output_137_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_137_cast_fp16 = matmul(transpose_x = attn_output_137_transpose_x_1, transpose_y = attn_output_137_transpose_y_1, x = attn_weights_287_cast_fp16, y = var_6558_cast_fp16_1)[name = string("attn_output_137_cast_fp16")];
            int32 var_6588 = const()[name = string("op_6588"), val = int32(1)];
            bool attn_output_139_interleave_0 = const()[name = string("attn_output_139_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_139_cast_fp16 = concat(axis = var_6588, interleave = attn_output_139_interleave_0, values = (var_6574_cast_fp16, attn_output_137_cast_fp16))[name = string("attn_output_139_cast_fp16")];
            tensor<int32, [4]> var_6592_perm_0 = const()[name = string("op_6592_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_6597 = const()[name = string("op_6597"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_6592_cast_fp16 = transpose(perm = var_6592_perm_0, x = attn_output_139_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_143_cast_fp16 = reshape(shape = var_6597, x = var_6592_cast_fp16)[name = string("attn_output_143_cast_fp16")];
            string hidden_states_173_pad_type_0 = const()[name = string("hidden_states_173_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_173_strides_0 = const()[name = string("hidden_states_173_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_173_pad_0 = const()[name = string("hidden_states_173_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_173_dilations_0 = const()[name = string("hidden_states_173_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_173_groups_0 = const()[name = string("hidden_states_173_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_173_cast_fp16 = conv(dilations = hidden_states_173_dilations_0, groups = hidden_states_173_groups_0, pad = hidden_states_173_pad_0, pad_type = hidden_states_173_pad_type_0, strides = hidden_states_173_strides_0, weight = layers_17_self_attn_o_proj_weight_cast_fp16, x = attn_output_143_cast_fp16)[name = string("hidden_states_173_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_175_cast_fp16 = add(x = hidden_states_169_cast_fp16, y = hidden_states_173_cast_fp16)[name = string("hidden_states_175_cast_fp16")];
            int32 var_6623 = const()[name = string("op_6623"), val = int32(1)];
            fp16 const_217_promoted_to_fp16 = const()[name = string("const_217_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_6625_cast_fp16 = mul(x = hidden_states_175_cast_fp16, y = const_217_promoted_to_fp16)[name = string("op_6625_cast_fp16")];
            bool doubled_141_interleave_0 = const()[name = string("doubled_141_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_141_cast_fp16 = concat(axis = var_6623, interleave = doubled_141_interleave_0, values = (hidden_states_175_cast_fp16, var_6625_cast_fp16))[name = string("doubled_141_cast_fp16")];
            tensor<int32, [1]> out_71_axes_0 = const()[name = string("out_71_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_71_gamma_0_to_fp16 = const()[name = string("out_71_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1483664192)))];
            fp16 var_6635_to_fp16 = const()[name = string("op_6635_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_6635_to_fp16, gamma = out_71_gamma_0_to_fp16, x = doubled_141_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<int32, [2]> var_6646_split_sizes_0 = const()[name = string("op_6646_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_6646_axis_0 = const()[name = string("op_6646_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_6646_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_6646_cast_fp16_1 = split(axis = var_6646_axis_0, split_sizes = var_6646_split_sizes_0, x = out_71_cast_fp16)[name = string("op_6646_cast_fp16")];
            string input_35_pad_type_0 = const()[name = string("input_35_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_35_strides_0 = const()[name = string("input_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_35_pad_0 = const()[name = string("input_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_35_dilations_0 = const()[name = string("input_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_35_groups_0 = const()[name = string("input_35_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_35_cast_fp16 = conv(dilations = input_35_dilations_0, groups = input_35_groups_0, pad = input_35_pad_0, pad_type = input_35_pad_type_0, strides = input_35_strides_0, weight = layers_17_mlp_gate_proj_weight_cast_fp16, x = var_6646_cast_fp16_0)[name = string("input_35_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_6663_cast_fp16 = silu(x = input_35_cast_fp16)[name = string("op_6663_cast_fp16")];
            string var_6669_pad_type_0 = const()[name = string("op_6669_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6669_strides_0 = const()[name = string("op_6669_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6669_pad_0 = const()[name = string("op_6669_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6669_dilations_0 = const()[name = string("op_6669_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6669_groups_0 = const()[name = string("op_6669_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_6669_cast_fp16 = conv(dilations = var_6669_dilations_0, groups = var_6669_groups_0, pad = var_6669_pad_0, pad_type = var_6669_pad_type_0, strides = var_6669_strides_0, weight = layers_17_mlp_up_proj_weight_cast_fp16, x = var_6646_cast_fp16_0)[name = string("op_6669_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_179_cast_fp16 = mul(x = var_6663_cast_fp16, y = var_6669_cast_fp16)[name = string("x_179_cast_fp16")];
            string hidden_states_177_pad_type_0 = const()[name = string("hidden_states_177_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_177_strides_0 = const()[name = string("hidden_states_177_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_177_pad_0 = const()[name = string("hidden_states_177_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_177_dilations_0 = const()[name = string("hidden_states_177_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_177_groups_0 = const()[name = string("hidden_states_177_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_177_cast_fp16 = conv(dilations = hidden_states_177_dilations_0, groups = hidden_states_177_groups_0, pad = hidden_states_177_pad_0, pad_type = hidden_states_177_pad_type_0, strides = hidden_states_177_strides_0, weight = layers_17_mlp_down_proj_weight_cast_fp16, x = x_179_cast_fp16)[name = string("hidden_states_177_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_179_cast_fp16 = add(x = hidden_states_175_cast_fp16, y = hidden_states_177_cast_fp16)[name = string("hidden_states_179_cast_fp16")];
            int32 var_6685 = const()[name = string("op_6685"), val = int32(1)];
            fp16 const_219_promoted_to_fp16 = const()[name = string("const_219_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_6687_cast_fp16 = mul(x = hidden_states_179_cast_fp16, y = const_219_promoted_to_fp16)[name = string("op_6687_cast_fp16")];
            bool doubled_145_interleave_0 = const()[name = string("doubled_145_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_145_cast_fp16 = concat(axis = var_6685, interleave = doubled_145_interleave_0, values = (hidden_states_179_cast_fp16, var_6687_cast_fp16))[name = string("doubled_145_cast_fp16")];
            tensor<int32, [1]> out_73_axes_0 = const()[name = string("out_73_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_73_gamma_0_to_fp16 = const()[name = string("out_73_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1483672448)))];
            fp16 var_6697_to_fp16 = const()[name = string("op_6697_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_6697_to_fp16, gamma = out_73_gamma_0_to_fp16, x = doubled_145_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<int32, [2]> var_6708_split_sizes_0 = const()[name = string("op_6708_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_6708_axis_0 = const()[name = string("op_6708_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_6708_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_6708_cast_fp16_1 = split(axis = var_6708_axis_0, split_sizes = var_6708_split_sizes_0, x = out_73_cast_fp16)[name = string("op_6708_cast_fp16")];
            string query_states_109_pad_type_0 = const()[name = string("query_states_109_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_109_strides_0 = const()[name = string("query_states_109_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_109_pad_0 = const()[name = string("query_states_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_109_dilations_0 = const()[name = string("query_states_109_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_109_groups_0 = const()[name = string("query_states_109_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_109_cast_fp16 = conv(dilations = query_states_109_dilations_0, groups = query_states_109_groups_0, pad = query_states_109_pad_0, pad_type = query_states_109_pad_type_0, strides = query_states_109_strides_0, weight = layers_18_self_attn_q_proj_weight_cast_fp16, x = var_6708_cast_fp16_0)[name = string("query_states_109_cast_fp16")];
            string key_states_181_pad_type_0 = const()[name = string("key_states_181_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_181_strides_0 = const()[name = string("key_states_181_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_181_pad_0 = const()[name = string("key_states_181_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_181_dilations_0 = const()[name = string("key_states_181_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_181_groups_0 = const()[name = string("key_states_181_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_181_cast_fp16 = conv(dilations = key_states_181_dilations_0, groups = key_states_181_groups_0, pad = key_states_181_pad_0, pad_type = key_states_181_pad_type_0, strides = key_states_181_strides_0, weight = layers_18_self_attn_k_proj_weight_cast_fp16, x = var_6708_cast_fp16_0)[name = string("key_states_181_cast_fp16")];
            string value_states_109_pad_type_0 = const()[name = string("value_states_109_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_109_strides_0 = const()[name = string("value_states_109_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_109_pad_0 = const()[name = string("value_states_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_109_dilations_0 = const()[name = string("value_states_109_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_109_groups_0 = const()[name = string("value_states_109_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> value_states_109_cast_fp16 = conv(dilations = value_states_109_dilations_0, groups = value_states_109_groups_0, pad = value_states_109_pad_0, pad_type = value_states_109_pad_type_0, strides = value_states_109_strides_0, weight = layers_18_self_attn_v_proj_weight_cast_fp16, x = var_6708_cast_fp16_0)[name = string("value_states_109_cast_fp16")];
            tensor<int32, [4]> var_6760 = const()[name = string("op_6760"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_181_cast_fp16 = reshape(shape = var_6760, x = query_states_109_cast_fp16)[name = string("x_181_cast_fp16")];
            tensor<int32, [4]> var_6764 = const()[name = string("op_6764"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_6765_cast_fp16 = reshape(shape = var_6764, x = key_states_181_cast_fp16)[name = string("op_6765_cast_fp16")];
            tensor<int32, [4]> var_6771 = const()[name = string("op_6771"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_6772_cast_fp16 = reshape(shape = var_6771, x = value_states_109_cast_fp16)[name = string("op_6772_cast_fp16")];
            tensor<int32, [4]> value_states_111_perm_0 = const()[name = string("value_states_111_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_6776_cast_fp16 = mul(x = x_181_cast_fp16, y = var_859_cast_fp16)[name = string("op_6776_cast_fp16")];
            tensor<int32, [2]> var_6777_split_sizes_0 = const()[name = string("op_6777_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_6777_axis_0 = const()[name = string("op_6777_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_6777_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_6777_cast_fp16_1 = split(axis = var_6777_axis_0, split_sizes = var_6777_split_sizes_0, x = x_181_cast_fp16)[name = string("op_6777_cast_fp16")];
            fp16 const_223_promoted_to_fp16 = const()[name = string("const_223_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_6779_cast_fp16 = mul(x = var_6777_cast_fp16_1, y = const_223_promoted_to_fp16)[name = string("op_6779_cast_fp16")];
            int32 var_6781 = const()[name = string("op_6781"), val = int32(-2)];
            bool var_6782_interleave_0 = const()[name = string("op_6782_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_6782_cast_fp16 = concat(axis = var_6781, interleave = var_6782_interleave_0, values = (var_6779_cast_fp16, var_6777_cast_fp16_0))[name = string("op_6782_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_6783_cast_fp16 = mul(x = var_6782_cast_fp16, y = var_866_cast_fp16)[name = string("op_6783_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_111_cast_fp16 = add(x = var_6776_cast_fp16, y = var_6783_cast_fp16)[name = string("query_states_111_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_6789_cast_fp16 = mul(x = var_6765_cast_fp16, y = var_859_cast_fp16)[name = string("op_6789_cast_fp16")];
            tensor<int32, [2]> var_6790_split_sizes_0 = const()[name = string("op_6790_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_6790_axis_0 = const()[name = string("op_6790_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_6790_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_6790_cast_fp16_1 = split(axis = var_6790_axis_0, split_sizes = var_6790_split_sizes_0, x = var_6765_cast_fp16)[name = string("op_6790_cast_fp16")];
            fp16 const_224_promoted_to_fp16 = const()[name = string("const_224_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_6792_cast_fp16 = mul(x = var_6790_cast_fp16_1, y = const_224_promoted_to_fp16)[name = string("op_6792_cast_fp16")];
            int32 var_6794 = const()[name = string("op_6794"), val = int32(-2)];
            bool var_6795_interleave_0 = const()[name = string("op_6795_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_6795_cast_fp16 = concat(axis = var_6794, interleave = var_6795_interleave_0, values = (var_6792_cast_fp16, var_6790_cast_fp16_0))[name = string("op_6795_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_6796_cast_fp16 = mul(x = var_6795_cast_fp16, y = var_866_cast_fp16)[name = string("op_6796_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_185_cast_fp16 = add(x = var_6789_cast_fp16, y = var_6796_cast_fp16)[name = string("key_states_185_cast_fp16")];
            tensor<int32, [4]> key_states_187_perm_0 = const()[name = string("key_states_187_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_217 = const()[name = string("expand_dims_217"), val = tensor<int32, [1]>([18])];
            tensor<int32, [1]> expand_dims_218 = const()[name = string("expand_dims_218"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_220 = const()[name = string("expand_dims_220"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_221 = const()[name = string("expand_dims_221"), val = tensor<int32, [1]>([19])];
            int32 concat_147_axis_0 = const()[name = string("concat_147_axis_0"), val = int32(0)];
            bool concat_147_interleave_0 = const()[name = string("concat_147_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_147 = concat(axis = concat_147_axis_0, interleave = concat_147_interleave_0, values = (expand_dims_217, expand_dims_218, position_id, expand_dims_220))[name = string("concat_147")];
            tensor<int32, [1]> concat_148_values1_0 = const()[name = string("concat_148_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_148_values3_0 = const()[name = string("concat_148_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_148_axis_0 = const()[name = string("concat_148_axis_0"), val = int32(0)];
            bool concat_148_interleave_0 = const()[name = string("concat_148_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_148 = concat(axis = concat_148_axis_0, interleave = concat_148_interleave_0, values = (expand_dims_221, concat_148_values1_0, cache_position_end, concat_148_values3_0))[name = string("concat_148")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_19_stride_0 = const()[name = string("key_cache_internal_tensor_assign_19_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_19_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_19_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_19_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_19_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_19_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_19_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_187_cast_fp16 = transpose(perm = key_states_187_perm_0, x = key_states_185_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_19_cast_fp16 = slice_update(begin = concat_147, begin_mask = key_cache_internal_tensor_assign_19_begin_mask_0, end = concat_148, end_mask = key_cache_internal_tensor_assign_19_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_19_squeeze_mask_0, stride = key_cache_internal_tensor_assign_19_stride_0, update = key_states_187_cast_fp16, x = coreml_update_state_90)[name = string("key_cache_internal_tensor_assign_19_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_19_cast_fp16, input = key_cache)[name = string("coreml_update_state_92_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_92 = read_state(input = key_cache)[name = string("coreml_update_state_92")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_19_stride_0 = const()[name = string("value_cache_internal_tensor_assign_19_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_19_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_19_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_19_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_19_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_19_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_19_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_111_cast_fp16 = transpose(perm = value_states_111_perm_0, x = var_6772_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_19_cast_fp16 = slice_update(begin = concat_147, begin_mask = value_cache_internal_tensor_assign_19_begin_mask_0, end = concat_148, end_mask = value_cache_internal_tensor_assign_19_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_19_squeeze_mask_0, stride = value_cache_internal_tensor_assign_19_stride_0, update = value_states_111_cast_fp16, x = coreml_update_state_91)[name = string("value_cache_internal_tensor_assign_19_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_19_cast_fp16, input = value_cache)[name = string("coreml_update_state_93_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_93 = read_state(input = value_cache)[name = string("coreml_update_state_93")];
            tensor<int32, [4]> var_6866_begin_0 = const()[name = string("op_6866_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_6866_end_0 = const()[name = string("op_6866_end_0"), val = tensor<int32, [4]>([19, 2, 2048, 128])];
            tensor<bool, [4]> var_6866_end_mask_0 = const()[name = string("op_6866_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_6866_cast_fp16 = slice_by_index(begin = var_6866_begin_0, end = var_6866_end_0, end_mask = var_6866_end_mask_0, x = coreml_update_state_92)[name = string("op_6866_cast_fp16")];
            tensor<int32, [2]> tile_36 = const()[name = string("tile_36"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6869_axis_0 = const()[name = string("op_6869_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_6869_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_6869_cast_fp16_1 = split(axis = var_6869_axis_0, split_sizes = tile_36, x = var_6866_cast_fp16)[name = string("op_6869_cast_fp16")];
            tensor<int32, [4]> var_6876_begin_0 = const()[name = string("op_6876_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_6876_end_0 = const()[name = string("op_6876_end_0"), val = tensor<int32, [4]>([19, 2, 2048, 128])];
            tensor<bool, [4]> var_6876_end_mask_0 = const()[name = string("op_6876_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_6876_cast_fp16 = slice_by_index(begin = var_6876_begin_0, end = var_6876_end_0, end_mask = var_6876_end_mask_0, x = coreml_update_state_93)[name = string("op_6876_cast_fp16")];
            tensor<int32, [2]> tile_37 = const()[name = string("tile_37"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6879_axis_0 = const()[name = string("op_6879_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_6879_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_6879_cast_fp16_1 = split(axis = var_6879_axis_0, split_sizes = tile_37, x = var_6876_cast_fp16)[name = string("op_6879_cast_fp16")];
            tensor<int32, [2]> var_6882_split_sizes_0 = const()[name = string("op_6882_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_6882_axis_0 = const()[name = string("op_6882_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_6882_0, tensor<fp16, [1, 8, 128, 8]> var_6882_1 = split(axis = var_6882_axis_0, split_sizes = var_6882_split_sizes_0, x = query_states_111_cast_fp16)[name = string("op_6882")];
            bool attn_weights_289_transpose_x_0 = const()[name = string("attn_weights_289_transpose_x_0"), val = bool(false)];
            bool attn_weights_289_transpose_y_0 = const()[name = string("attn_weights_289_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_289_cast_fp16 = matmul(transpose_x = attn_weights_289_transpose_x_0, transpose_y = attn_weights_289_transpose_y_0, x = var_6869_cast_fp16_0, y = var_6882_0)[name = string("attn_weights_289_cast_fp16")];
            fp16 var_6885_to_fp16 = const()[name = string("op_6885_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_291_cast_fp16 = mul(x = attn_weights_289_cast_fp16, y = var_6885_to_fp16)[name = string("attn_weights_291_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_293_cast_fp16 = add(x = attn_weights_291_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_293_cast_fp16")];
            int32 var_6889 = const()[name = string("op_6889"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_295_cast_fp16 = softmax(axis = var_6889, x = attn_weights_293_cast_fp16)[name = string("attn_weights_295_cast_fp16")];
            bool var_6895_transpose_x_1 = const()[name = string("op_6895_transpose_x_1"), val = bool(true)];
            bool var_6895_transpose_y_1 = const()[name = string("op_6895_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_6895_cast_fp16 = matmul(transpose_x = var_6895_transpose_x_1, transpose_y = var_6895_transpose_y_1, x = attn_weights_295_cast_fp16, y = var_6879_cast_fp16_0)[name = string("op_6895_cast_fp16")];
            bool attn_weights_297_transpose_x_0 = const()[name = string("attn_weights_297_transpose_x_0"), val = bool(false)];
            bool attn_weights_297_transpose_y_0 = const()[name = string("attn_weights_297_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_297_cast_fp16 = matmul(transpose_x = attn_weights_297_transpose_x_0, transpose_y = attn_weights_297_transpose_y_0, x = var_6869_cast_fp16_1, y = var_6882_1)[name = string("attn_weights_297_cast_fp16")];
            fp16 var_6897_to_fp16 = const()[name = string("op_6897_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_299_cast_fp16 = mul(x = attn_weights_297_cast_fp16, y = var_6897_to_fp16)[name = string("attn_weights_299_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_301_cast_fp16 = add(x = attn_weights_299_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_301_cast_fp16")];
            int32 var_6901 = const()[name = string("op_6901"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_303_cast_fp16 = softmax(axis = var_6901, x = attn_weights_301_cast_fp16)[name = string("attn_weights_303_cast_fp16")];
            bool attn_output_145_transpose_x_1 = const()[name = string("attn_output_145_transpose_x_1"), val = bool(true)];
            bool attn_output_145_transpose_y_1 = const()[name = string("attn_output_145_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_145_cast_fp16 = matmul(transpose_x = attn_output_145_transpose_x_1, transpose_y = attn_output_145_transpose_y_1, x = attn_weights_303_cast_fp16, y = var_6879_cast_fp16_1)[name = string("attn_output_145_cast_fp16")];
            int32 var_6909 = const()[name = string("op_6909"), val = int32(1)];
            bool attn_output_147_interleave_0 = const()[name = string("attn_output_147_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_147_cast_fp16 = concat(axis = var_6909, interleave = attn_output_147_interleave_0, values = (var_6895_cast_fp16, attn_output_145_cast_fp16))[name = string("attn_output_147_cast_fp16")];
            tensor<int32, [4]> var_6913_perm_0 = const()[name = string("op_6913_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_6918 = const()[name = string("op_6918"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_6913_cast_fp16 = transpose(perm = var_6913_perm_0, x = attn_output_147_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_151_cast_fp16 = reshape(shape = var_6918, x = var_6913_cast_fp16)[name = string("attn_output_151_cast_fp16")];
            string hidden_states_183_pad_type_0 = const()[name = string("hidden_states_183_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_183_strides_0 = const()[name = string("hidden_states_183_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_183_pad_0 = const()[name = string("hidden_states_183_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_183_dilations_0 = const()[name = string("hidden_states_183_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_183_groups_0 = const()[name = string("hidden_states_183_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_183_cast_fp16 = conv(dilations = hidden_states_183_dilations_0, groups = hidden_states_183_groups_0, pad = hidden_states_183_pad_0, pad_type = hidden_states_183_pad_type_0, strides = hidden_states_183_strides_0, weight = layers_18_self_attn_o_proj_weight_cast_fp16, x = attn_output_151_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_185_cast_fp16 = add(x = hidden_states_179_cast_fp16, y = hidden_states_183_cast_fp16)[name = string("hidden_states_185_cast_fp16")];
            int32 var_6944 = const()[name = string("op_6944"), val = int32(1)];
            fp16 const_229_promoted_to_fp16 = const()[name = string("const_229_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_6946_cast_fp16 = mul(x = hidden_states_185_cast_fp16, y = const_229_promoted_to_fp16)[name = string("op_6946_cast_fp16")];
            bool doubled_149_interleave_0 = const()[name = string("doubled_149_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_149_cast_fp16 = concat(axis = var_6944, interleave = doubled_149_interleave_0, values = (hidden_states_185_cast_fp16, var_6946_cast_fp16))[name = string("doubled_149_cast_fp16")];
            tensor<int32, [1]> out_75_axes_0 = const()[name = string("out_75_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_75_gamma_0_to_fp16 = const()[name = string("out_75_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1483680704)))];
            fp16 var_6956_to_fp16 = const()[name = string("op_6956_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_75_cast_fp16 = layer_norm(axes = out_75_axes_0, epsilon = var_6956_to_fp16, gamma = out_75_gamma_0_to_fp16, x = doubled_149_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<int32, [2]> var_6967_split_sizes_0 = const()[name = string("op_6967_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_6967_axis_0 = const()[name = string("op_6967_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_6967_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_6967_cast_fp16_1 = split(axis = var_6967_axis_0, split_sizes = var_6967_split_sizes_0, x = out_75_cast_fp16)[name = string("op_6967_cast_fp16")];
            string input_37_pad_type_0 = const()[name = string("input_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_37_strides_0 = const()[name = string("input_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_37_pad_0 = const()[name = string("input_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_37_dilations_0 = const()[name = string("input_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_37_groups_0 = const()[name = string("input_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_37_cast_fp16 = conv(dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_18_mlp_gate_proj_weight_cast_fp16, x = var_6967_cast_fp16_0)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_6984_cast_fp16 = silu(x = input_37_cast_fp16)[name = string("op_6984_cast_fp16")];
            string var_6990_pad_type_0 = const()[name = string("op_6990_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6990_strides_0 = const()[name = string("op_6990_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6990_pad_0 = const()[name = string("op_6990_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6990_dilations_0 = const()[name = string("op_6990_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6990_groups_0 = const()[name = string("op_6990_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_6990_cast_fp16 = conv(dilations = var_6990_dilations_0, groups = var_6990_groups_0, pad = var_6990_pad_0, pad_type = var_6990_pad_type_0, strides = var_6990_strides_0, weight = layers_18_mlp_up_proj_weight_cast_fp16, x = var_6967_cast_fp16_0)[name = string("op_6990_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_189_cast_fp16 = mul(x = var_6984_cast_fp16, y = var_6990_cast_fp16)[name = string("x_189_cast_fp16")];
            string hidden_states_187_pad_type_0 = const()[name = string("hidden_states_187_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_187_strides_0 = const()[name = string("hidden_states_187_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_187_pad_0 = const()[name = string("hidden_states_187_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_187_dilations_0 = const()[name = string("hidden_states_187_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_187_groups_0 = const()[name = string("hidden_states_187_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_187_cast_fp16 = conv(dilations = hidden_states_187_dilations_0, groups = hidden_states_187_groups_0, pad = hidden_states_187_pad_0, pad_type = hidden_states_187_pad_type_0, strides = hidden_states_187_strides_0, weight = layers_18_mlp_down_proj_weight_cast_fp16, x = x_189_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_189_cast_fp16 = add(x = hidden_states_185_cast_fp16, y = hidden_states_187_cast_fp16)[name = string("hidden_states_189_cast_fp16")];
            int32 var_7006 = const()[name = string("op_7006"), val = int32(1)];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_7008_cast_fp16 = mul(x = hidden_states_189_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_7008_cast_fp16")];
            bool doubled_153_interleave_0 = const()[name = string("doubled_153_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_153_cast_fp16 = concat(axis = var_7006, interleave = doubled_153_interleave_0, values = (hidden_states_189_cast_fp16, var_7008_cast_fp16))[name = string("doubled_153_cast_fp16")];
            tensor<int32, [1]> out_77_axes_0 = const()[name = string("out_77_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_77_gamma_0_to_fp16 = const()[name = string("out_77_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1483688960)))];
            fp16 var_7018_to_fp16 = const()[name = string("op_7018_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_77_cast_fp16 = layer_norm(axes = out_77_axes_0, epsilon = var_7018_to_fp16, gamma = out_77_gamma_0_to_fp16, x = doubled_153_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<int32, [2]> var_7029_split_sizes_0 = const()[name = string("op_7029_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_7029_axis_0 = const()[name = string("op_7029_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_7029_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_7029_cast_fp16_1 = split(axis = var_7029_axis_0, split_sizes = var_7029_split_sizes_0, x = out_77_cast_fp16)[name = string("op_7029_cast_fp16")];
            string query_states_115_pad_type_0 = const()[name = string("query_states_115_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_115_strides_0 = const()[name = string("query_states_115_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_115_pad_0 = const()[name = string("query_states_115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_115_dilations_0 = const()[name = string("query_states_115_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_115_groups_0 = const()[name = string("query_states_115_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_115_cast_fp16 = conv(dilations = query_states_115_dilations_0, groups = query_states_115_groups_0, pad = query_states_115_pad_0, pad_type = query_states_115_pad_type_0, strides = query_states_115_strides_0, weight = layers_19_self_attn_q_proj_weight_cast_fp16, x = var_7029_cast_fp16_0)[name = string("query_states_115_cast_fp16")];
            string key_states_191_pad_type_0 = const()[name = string("key_states_191_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_191_strides_0 = const()[name = string("key_states_191_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_191_pad_0 = const()[name = string("key_states_191_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_191_dilations_0 = const()[name = string("key_states_191_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_191_groups_0 = const()[name = string("key_states_191_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_191_cast_fp16 = conv(dilations = key_states_191_dilations_0, groups = key_states_191_groups_0, pad = key_states_191_pad_0, pad_type = key_states_191_pad_type_0, strides = key_states_191_strides_0, weight = layers_19_self_attn_k_proj_weight_cast_fp16, x = var_7029_cast_fp16_0)[name = string("key_states_191_cast_fp16")];
            string value_states_115_pad_type_0 = const()[name = string("value_states_115_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_115_strides_0 = const()[name = string("value_states_115_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_115_pad_0 = const()[name = string("value_states_115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_115_dilations_0 = const()[name = string("value_states_115_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_115_groups_0 = const()[name = string("value_states_115_groups_0"), val = int32(1)];
            tensor<fp16, [256, 2048, 1, 1]> layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1483697216)))];
            tensor<fp16, [1, 256, 1, 8]> value_states_115_cast_fp16 = conv(dilations = value_states_115_dilations_0, groups = value_states_115_groups_0, pad = value_states_115_pad_0, pad_type = value_states_115_pad_type_0, strides = value_states_115_strides_0, weight = layers_19_self_attn_v_proj_weight_to_fp16, x = var_7029_cast_fp16_0)[name = string("value_states_115_cast_fp16")];
            tensor<int32, [4]> var_7081 = const()[name = string("op_7081"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_191_cast_fp16 = reshape(shape = var_7081, x = query_states_115_cast_fp16)[name = string("x_191_cast_fp16")];
            tensor<int32, [4]> var_7085 = const()[name = string("op_7085"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_7086_cast_fp16 = reshape(shape = var_7085, x = key_states_191_cast_fp16)[name = string("op_7086_cast_fp16")];
            tensor<int32, [4]> var_7092 = const()[name = string("op_7092"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_7093_cast_fp16 = reshape(shape = var_7092, x = value_states_115_cast_fp16)[name = string("op_7093_cast_fp16")];
            tensor<int32, [4]> value_states_117_perm_0 = const()[name = string("value_states_117_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_7097_cast_fp16 = mul(x = x_191_cast_fp16, y = var_859_cast_fp16)[name = string("op_7097_cast_fp16")];
            tensor<int32, [2]> var_7098_split_sizes_0 = const()[name = string("op_7098_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_7098_axis_0 = const()[name = string("op_7098_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_7098_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_7098_cast_fp16_1 = split(axis = var_7098_axis_0, split_sizes = var_7098_split_sizes_0, x = x_191_cast_fp16)[name = string("op_7098_cast_fp16")];
            fp16 const_235_promoted_to_fp16 = const()[name = string("const_235_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_7100_cast_fp16 = mul(x = var_7098_cast_fp16_1, y = const_235_promoted_to_fp16)[name = string("op_7100_cast_fp16")];
            int32 var_7102 = const()[name = string("op_7102"), val = int32(-2)];
            bool var_7103_interleave_0 = const()[name = string("op_7103_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_7103_cast_fp16 = concat(axis = var_7102, interleave = var_7103_interleave_0, values = (var_7100_cast_fp16, var_7098_cast_fp16_0))[name = string("op_7103_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_7104_cast_fp16 = mul(x = var_7103_cast_fp16, y = var_866_cast_fp16)[name = string("op_7104_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_117_cast_fp16 = add(x = var_7097_cast_fp16, y = var_7104_cast_fp16)[name = string("query_states_117_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_7110_cast_fp16 = mul(x = var_7086_cast_fp16, y = var_859_cast_fp16)[name = string("op_7110_cast_fp16")];
            tensor<int32, [2]> var_7111_split_sizes_0 = const()[name = string("op_7111_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_7111_axis_0 = const()[name = string("op_7111_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_7111_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_7111_cast_fp16_1 = split(axis = var_7111_axis_0, split_sizes = var_7111_split_sizes_0, x = var_7086_cast_fp16)[name = string("op_7111_cast_fp16")];
            fp16 const_236_promoted_to_fp16 = const()[name = string("const_236_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_7113_cast_fp16 = mul(x = var_7111_cast_fp16_1, y = const_236_promoted_to_fp16)[name = string("op_7113_cast_fp16")];
            int32 var_7115 = const()[name = string("op_7115"), val = int32(-2)];
            bool var_7116_interleave_0 = const()[name = string("op_7116_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_7116_cast_fp16 = concat(axis = var_7115, interleave = var_7116_interleave_0, values = (var_7113_cast_fp16, var_7111_cast_fp16_0))[name = string("op_7116_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_7117_cast_fp16 = mul(x = var_7116_cast_fp16, y = var_866_cast_fp16)[name = string("op_7117_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_195_cast_fp16 = add(x = var_7110_cast_fp16, y = var_7117_cast_fp16)[name = string("key_states_195_cast_fp16")];
            tensor<int32, [4]> key_states_197_perm_0 = const()[name = string("key_states_197_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_229 = const()[name = string("expand_dims_229"), val = tensor<int32, [1]>([19])];
            tensor<int32, [1]> expand_dims_230 = const()[name = string("expand_dims_230"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_232 = const()[name = string("expand_dims_232"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_233 = const()[name = string("expand_dims_233"), val = tensor<int32, [1]>([20])];
            int32 concat_155_axis_0 = const()[name = string("concat_155_axis_0"), val = int32(0)];
            bool concat_155_interleave_0 = const()[name = string("concat_155_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_155 = concat(axis = concat_155_axis_0, interleave = concat_155_interleave_0, values = (expand_dims_229, expand_dims_230, position_id, expand_dims_232))[name = string("concat_155")];
            tensor<int32, [1]> concat_156_values1_0 = const()[name = string("concat_156_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_156_values3_0 = const()[name = string("concat_156_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_156_axis_0 = const()[name = string("concat_156_axis_0"), val = int32(0)];
            bool concat_156_interleave_0 = const()[name = string("concat_156_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_156 = concat(axis = concat_156_axis_0, interleave = concat_156_interleave_0, values = (expand_dims_233, concat_156_values1_0, cache_position_end, concat_156_values3_0))[name = string("concat_156")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_20_stride_0 = const()[name = string("key_cache_internal_tensor_assign_20_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_20_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_20_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_20_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_20_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_20_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_20_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_197_cast_fp16 = transpose(perm = key_states_197_perm_0, x = key_states_195_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_20_cast_fp16 = slice_update(begin = concat_155, begin_mask = key_cache_internal_tensor_assign_20_begin_mask_0, end = concat_156, end_mask = key_cache_internal_tensor_assign_20_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_20_squeeze_mask_0, stride = key_cache_internal_tensor_assign_20_stride_0, update = key_states_197_cast_fp16, x = coreml_update_state_92)[name = string("key_cache_internal_tensor_assign_20_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_20_cast_fp16, input = key_cache)[name = string("coreml_update_state_94_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_94 = read_state(input = key_cache)[name = string("coreml_update_state_94")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_20_stride_0 = const()[name = string("value_cache_internal_tensor_assign_20_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_20_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_20_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_20_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_20_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_20_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_20_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_117_cast_fp16 = transpose(perm = value_states_117_perm_0, x = var_7093_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_20_cast_fp16 = slice_update(begin = concat_155, begin_mask = value_cache_internal_tensor_assign_20_begin_mask_0, end = concat_156, end_mask = value_cache_internal_tensor_assign_20_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_20_squeeze_mask_0, stride = value_cache_internal_tensor_assign_20_stride_0, update = value_states_117_cast_fp16, x = coreml_update_state_93)[name = string("value_cache_internal_tensor_assign_20_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_20_cast_fp16, input = value_cache)[name = string("coreml_update_state_95_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_95 = read_state(input = value_cache)[name = string("coreml_update_state_95")];
            tensor<int32, [4]> var_7187_begin_0 = const()[name = string("op_7187_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7187_end_0 = const()[name = string("op_7187_end_0"), val = tensor<int32, [4]>([20, 2, 2048, 128])];
            tensor<bool, [4]> var_7187_end_mask_0 = const()[name = string("op_7187_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_7187_cast_fp16 = slice_by_index(begin = var_7187_begin_0, end = var_7187_end_0, end_mask = var_7187_end_mask_0, x = coreml_update_state_94)[name = string("op_7187_cast_fp16")];
            tensor<int32, [2]> tile_38 = const()[name = string("tile_38"), val = tensor<int32, [2]>([1, 1])];
            int32 var_7190_axis_0 = const()[name = string("op_7190_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_7190_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_7190_cast_fp16_1 = split(axis = var_7190_axis_0, split_sizes = tile_38, x = var_7187_cast_fp16)[name = string("op_7190_cast_fp16")];
            tensor<int32, [4]> var_7197_begin_0 = const()[name = string("op_7197_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7197_end_0 = const()[name = string("op_7197_end_0"), val = tensor<int32, [4]>([20, 2, 2048, 128])];
            tensor<bool, [4]> var_7197_end_mask_0 = const()[name = string("op_7197_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_7197_cast_fp16 = slice_by_index(begin = var_7197_begin_0, end = var_7197_end_0, end_mask = var_7197_end_mask_0, x = coreml_update_state_95)[name = string("op_7197_cast_fp16")];
            tensor<int32, [2]> tile_39 = const()[name = string("tile_39"), val = tensor<int32, [2]>([1, 1])];
            int32 var_7200_axis_0 = const()[name = string("op_7200_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_7200_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_7200_cast_fp16_1 = split(axis = var_7200_axis_0, split_sizes = tile_39, x = var_7197_cast_fp16)[name = string("op_7200_cast_fp16")];
            tensor<int32, [2]> var_7203_split_sizes_0 = const()[name = string("op_7203_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_7203_axis_0 = const()[name = string("op_7203_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_7203_0, tensor<fp16, [1, 8, 128, 8]> var_7203_1 = split(axis = var_7203_axis_0, split_sizes = var_7203_split_sizes_0, x = query_states_117_cast_fp16)[name = string("op_7203")];
            bool attn_weights_305_transpose_x_0 = const()[name = string("attn_weights_305_transpose_x_0"), val = bool(false)];
            bool attn_weights_305_transpose_y_0 = const()[name = string("attn_weights_305_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_305_cast_fp16 = matmul(transpose_x = attn_weights_305_transpose_x_0, transpose_y = attn_weights_305_transpose_y_0, x = var_7190_cast_fp16_0, y = var_7203_0)[name = string("attn_weights_305_cast_fp16")];
            fp16 var_7206_to_fp16 = const()[name = string("op_7206_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_307_cast_fp16 = mul(x = attn_weights_305_cast_fp16, y = var_7206_to_fp16)[name = string("attn_weights_307_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_309_cast_fp16 = add(x = attn_weights_307_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_309_cast_fp16")];
            int32 var_7210 = const()[name = string("op_7210"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_311_cast_fp16 = softmax(axis = var_7210, x = attn_weights_309_cast_fp16)[name = string("attn_weights_311_cast_fp16")];
            bool var_7216_transpose_x_1 = const()[name = string("op_7216_transpose_x_1"), val = bool(true)];
            bool var_7216_transpose_y_1 = const()[name = string("op_7216_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_7216_cast_fp16 = matmul(transpose_x = var_7216_transpose_x_1, transpose_y = var_7216_transpose_y_1, x = attn_weights_311_cast_fp16, y = var_7200_cast_fp16_0)[name = string("op_7216_cast_fp16")];
            bool attn_weights_313_transpose_x_0 = const()[name = string("attn_weights_313_transpose_x_0"), val = bool(false)];
            bool attn_weights_313_transpose_y_0 = const()[name = string("attn_weights_313_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_313_cast_fp16 = matmul(transpose_x = attn_weights_313_transpose_x_0, transpose_y = attn_weights_313_transpose_y_0, x = var_7190_cast_fp16_1, y = var_7203_1)[name = string("attn_weights_313_cast_fp16")];
            fp16 var_7218_to_fp16 = const()[name = string("op_7218_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_315_cast_fp16 = mul(x = attn_weights_313_cast_fp16, y = var_7218_to_fp16)[name = string("attn_weights_315_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_317_cast_fp16 = add(x = attn_weights_315_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_317_cast_fp16")];
            int32 var_7222 = const()[name = string("op_7222"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_319_cast_fp16 = softmax(axis = var_7222, x = attn_weights_317_cast_fp16)[name = string("attn_weights_319_cast_fp16")];
            bool attn_output_153_transpose_x_1 = const()[name = string("attn_output_153_transpose_x_1"), val = bool(true)];
            bool attn_output_153_transpose_y_1 = const()[name = string("attn_output_153_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_153_cast_fp16 = matmul(transpose_x = attn_output_153_transpose_x_1, transpose_y = attn_output_153_transpose_y_1, x = attn_weights_319_cast_fp16, y = var_7200_cast_fp16_1)[name = string("attn_output_153_cast_fp16")];
            int32 var_7230 = const()[name = string("op_7230"), val = int32(1)];
            bool attn_output_155_interleave_0 = const()[name = string("attn_output_155_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_155_cast_fp16 = concat(axis = var_7230, interleave = attn_output_155_interleave_0, values = (var_7216_cast_fp16, attn_output_153_cast_fp16))[name = string("attn_output_155_cast_fp16")];
            tensor<int32, [4]> var_7234_perm_0 = const()[name = string("op_7234_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_7239 = const()[name = string("op_7239"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_7234_cast_fp16 = transpose(perm = var_7234_perm_0, x = attn_output_155_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_159_cast_fp16 = reshape(shape = var_7239, x = var_7234_cast_fp16)[name = string("attn_output_159_cast_fp16")];
            string hidden_states_193_pad_type_0 = const()[name = string("hidden_states_193_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_193_strides_0 = const()[name = string("hidden_states_193_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_193_pad_0 = const()[name = string("hidden_states_193_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_193_dilations_0 = const()[name = string("hidden_states_193_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_193_groups_0 = const()[name = string("hidden_states_193_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 2048, 1, 1]> layers_19_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_19_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1484745856)))];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_193_cast_fp16 = conv(dilations = hidden_states_193_dilations_0, groups = hidden_states_193_groups_0, pad = hidden_states_193_pad_0, pad_type = hidden_states_193_pad_type_0, strides = hidden_states_193_strides_0, weight = layers_19_self_attn_o_proj_weight_to_fp16, x = attn_output_159_cast_fp16)[name = string("hidden_states_193_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_195_cast_fp16 = add(x = hidden_states_189_cast_fp16, y = hidden_states_193_cast_fp16)[name = string("hidden_states_195_cast_fp16")];
            int32 var_7265 = const()[name = string("op_7265"), val = int32(1)];
            fp16 const_241_promoted_to_fp16 = const()[name = string("const_241_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_7267_cast_fp16 = mul(x = hidden_states_195_cast_fp16, y = const_241_promoted_to_fp16)[name = string("op_7267_cast_fp16")];
            bool doubled_157_interleave_0 = const()[name = string("doubled_157_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_157_cast_fp16 = concat(axis = var_7265, interleave = doubled_157_interleave_0, values = (hidden_states_195_cast_fp16, var_7267_cast_fp16))[name = string("doubled_157_cast_fp16")];
            tensor<int32, [1]> out_79_axes_0 = const()[name = string("out_79_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_79_gamma_0_to_fp16 = const()[name = string("out_79_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1493134528)))];
            fp16 var_7277_to_fp16 = const()[name = string("op_7277_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_7277_to_fp16, gamma = out_79_gamma_0_to_fp16, x = doubled_157_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<int32, [2]> var_7288_split_sizes_0 = const()[name = string("op_7288_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_7288_axis_0 = const()[name = string("op_7288_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_7288_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_7288_cast_fp16_1 = split(axis = var_7288_axis_0, split_sizes = var_7288_split_sizes_0, x = out_79_cast_fp16)[name = string("op_7288_cast_fp16")];
            string input_39_pad_type_0 = const()[name = string("input_39_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_39_strides_0 = const()[name = string("input_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_39_pad_0 = const()[name = string("input_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_39_dilations_0 = const()[name = string("input_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_39_groups_0 = const()[name = string("input_39_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_39_cast_fp16 = conv(dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = layers_19_mlp_gate_proj_weight_cast_fp16, x = var_7288_cast_fp16_0)[name = string("input_39_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_7305_cast_fp16 = silu(x = input_39_cast_fp16)[name = string("op_7305_cast_fp16")];
            string var_7311_pad_type_0 = const()[name = string("op_7311_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_7311_strides_0 = const()[name = string("op_7311_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7311_pad_0 = const()[name = string("op_7311_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7311_dilations_0 = const()[name = string("op_7311_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_7311_groups_0 = const()[name = string("op_7311_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_7311_cast_fp16 = conv(dilations = var_7311_dilations_0, groups = var_7311_groups_0, pad = var_7311_pad_0, pad_type = var_7311_pad_type_0, strides = var_7311_strides_0, weight = layers_19_mlp_up_proj_weight_cast_fp16, x = var_7288_cast_fp16_0)[name = string("op_7311_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_199_cast_fp16 = mul(x = var_7305_cast_fp16, y = var_7311_cast_fp16)[name = string("x_199_cast_fp16")];
            string hidden_states_197_pad_type_0 = const()[name = string("hidden_states_197_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_197_strides_0 = const()[name = string("hidden_states_197_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_197_pad_0 = const()[name = string("hidden_states_197_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_197_dilations_0 = const()[name = string("hidden_states_197_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_197_groups_0 = const()[name = string("hidden_states_197_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_197_cast_fp16 = conv(dilations = hidden_states_197_dilations_0, groups = hidden_states_197_groups_0, pad = hidden_states_197_pad_0, pad_type = hidden_states_197_pad_type_0, strides = hidden_states_197_strides_0, weight = layers_19_mlp_down_proj_weight_cast_fp16, x = x_199_cast_fp16)[name = string("hidden_states_197_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_199_cast_fp16 = add(x = hidden_states_195_cast_fp16, y = hidden_states_197_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            int32 var_7327 = const()[name = string("op_7327"), val = int32(1)];
            fp16 const_243_promoted_to_fp16 = const()[name = string("const_243_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_7329_cast_fp16 = mul(x = hidden_states_199_cast_fp16, y = const_243_promoted_to_fp16)[name = string("op_7329_cast_fp16")];
            bool doubled_161_interleave_0 = const()[name = string("doubled_161_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_161_cast_fp16 = concat(axis = var_7327, interleave = doubled_161_interleave_0, values = (hidden_states_199_cast_fp16, var_7329_cast_fp16))[name = string("doubled_161_cast_fp16")];
            tensor<int32, [1]> out_81_axes_0 = const()[name = string("out_81_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_81_gamma_0_to_fp16 = const()[name = string("out_81_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1493142784)))];
            fp16 var_7339_to_fp16 = const()[name = string("op_7339_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_81_cast_fp16 = layer_norm(axes = out_81_axes_0, epsilon = var_7339_to_fp16, gamma = out_81_gamma_0_to_fp16, x = doubled_161_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<int32, [2]> var_7350_split_sizes_0 = const()[name = string("op_7350_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_7350_axis_0 = const()[name = string("op_7350_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_7350_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_7350_cast_fp16_1 = split(axis = var_7350_axis_0, split_sizes = var_7350_split_sizes_0, x = out_81_cast_fp16)[name = string("op_7350_cast_fp16")];
            string query_states_121_pad_type_0 = const()[name = string("query_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_121_strides_0 = const()[name = string("query_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_121_pad_0 = const()[name = string("query_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_121_dilations_0 = const()[name = string("query_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_121_groups_0 = const()[name = string("query_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_121_cast_fp16 = conv(dilations = query_states_121_dilations_0, groups = query_states_121_groups_0, pad = query_states_121_pad_0, pad_type = query_states_121_pad_type_0, strides = query_states_121_strides_0, weight = layers_20_self_attn_q_proj_weight_cast_fp16, x = var_7350_cast_fp16_0)[name = string("query_states_121_cast_fp16")];
            string key_states_201_pad_type_0 = const()[name = string("key_states_201_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_201_strides_0 = const()[name = string("key_states_201_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_201_pad_0 = const()[name = string("key_states_201_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_201_dilations_0 = const()[name = string("key_states_201_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_201_groups_0 = const()[name = string("key_states_201_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_201_cast_fp16 = conv(dilations = key_states_201_dilations_0, groups = key_states_201_groups_0, pad = key_states_201_pad_0, pad_type = key_states_201_pad_type_0, strides = key_states_201_strides_0, weight = layers_20_self_attn_k_proj_weight_cast_fp16, x = var_7350_cast_fp16_0)[name = string("key_states_201_cast_fp16")];
            string value_states_121_pad_type_0 = const()[name = string("value_states_121_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_121_strides_0 = const()[name = string("value_states_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_121_pad_0 = const()[name = string("value_states_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_121_dilations_0 = const()[name = string("value_states_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_121_groups_0 = const()[name = string("value_states_121_groups_0"), val = int32(1)];
            tensor<fp16, [256, 2048, 1, 1]> layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1493151040)))];
            tensor<fp16, [1, 256, 1, 8]> value_states_121_cast_fp16 = conv(dilations = value_states_121_dilations_0, groups = value_states_121_groups_0, pad = value_states_121_pad_0, pad_type = value_states_121_pad_type_0, strides = value_states_121_strides_0, weight = layers_20_self_attn_v_proj_weight_to_fp16, x = var_7350_cast_fp16_0)[name = string("value_states_121_cast_fp16")];
            tensor<int32, [4]> var_7402 = const()[name = string("op_7402"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_201_cast_fp16 = reshape(shape = var_7402, x = query_states_121_cast_fp16)[name = string("x_201_cast_fp16")];
            tensor<int32, [4]> var_7406 = const()[name = string("op_7406"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_7407_cast_fp16 = reshape(shape = var_7406, x = key_states_201_cast_fp16)[name = string("op_7407_cast_fp16")];
            tensor<int32, [4]> var_7413 = const()[name = string("op_7413"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_7414_cast_fp16 = reshape(shape = var_7413, x = value_states_121_cast_fp16)[name = string("op_7414_cast_fp16")];
            tensor<int32, [4]> value_states_123_perm_0 = const()[name = string("value_states_123_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_7418_cast_fp16 = mul(x = x_201_cast_fp16, y = var_859_cast_fp16)[name = string("op_7418_cast_fp16")];
            tensor<int32, [2]> var_7419_split_sizes_0 = const()[name = string("op_7419_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_7419_axis_0 = const()[name = string("op_7419_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_7419_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_7419_cast_fp16_1 = split(axis = var_7419_axis_0, split_sizes = var_7419_split_sizes_0, x = x_201_cast_fp16)[name = string("op_7419_cast_fp16")];
            fp16 const_247_promoted_to_fp16 = const()[name = string("const_247_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_7421_cast_fp16 = mul(x = var_7419_cast_fp16_1, y = const_247_promoted_to_fp16)[name = string("op_7421_cast_fp16")];
            int32 var_7423 = const()[name = string("op_7423"), val = int32(-2)];
            bool var_7424_interleave_0 = const()[name = string("op_7424_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_7424_cast_fp16 = concat(axis = var_7423, interleave = var_7424_interleave_0, values = (var_7421_cast_fp16, var_7419_cast_fp16_0))[name = string("op_7424_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_7425_cast_fp16 = mul(x = var_7424_cast_fp16, y = var_866_cast_fp16)[name = string("op_7425_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_123_cast_fp16 = add(x = var_7418_cast_fp16, y = var_7425_cast_fp16)[name = string("query_states_123_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_7431_cast_fp16 = mul(x = var_7407_cast_fp16, y = var_859_cast_fp16)[name = string("op_7431_cast_fp16")];
            tensor<int32, [2]> var_7432_split_sizes_0 = const()[name = string("op_7432_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_7432_axis_0 = const()[name = string("op_7432_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_7432_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_7432_cast_fp16_1 = split(axis = var_7432_axis_0, split_sizes = var_7432_split_sizes_0, x = var_7407_cast_fp16)[name = string("op_7432_cast_fp16")];
            fp16 const_248_promoted_to_fp16 = const()[name = string("const_248_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_7434_cast_fp16 = mul(x = var_7432_cast_fp16_1, y = const_248_promoted_to_fp16)[name = string("op_7434_cast_fp16")];
            int32 var_7436 = const()[name = string("op_7436"), val = int32(-2)];
            bool var_7437_interleave_0 = const()[name = string("op_7437_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_7437_cast_fp16 = concat(axis = var_7436, interleave = var_7437_interleave_0, values = (var_7434_cast_fp16, var_7432_cast_fp16_0))[name = string("op_7437_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_7438_cast_fp16 = mul(x = var_7437_cast_fp16, y = var_866_cast_fp16)[name = string("op_7438_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_205_cast_fp16 = add(x = var_7431_cast_fp16, y = var_7438_cast_fp16)[name = string("key_states_205_cast_fp16")];
            tensor<int32, [4]> key_states_207_perm_0 = const()[name = string("key_states_207_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_241 = const()[name = string("expand_dims_241"), val = tensor<int32, [1]>([20])];
            tensor<int32, [1]> expand_dims_242 = const()[name = string("expand_dims_242"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_244 = const()[name = string("expand_dims_244"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_245 = const()[name = string("expand_dims_245"), val = tensor<int32, [1]>([21])];
            int32 concat_163_axis_0 = const()[name = string("concat_163_axis_0"), val = int32(0)];
            bool concat_163_interleave_0 = const()[name = string("concat_163_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_163 = concat(axis = concat_163_axis_0, interleave = concat_163_interleave_0, values = (expand_dims_241, expand_dims_242, position_id, expand_dims_244))[name = string("concat_163")];
            tensor<int32, [1]> concat_164_values1_0 = const()[name = string("concat_164_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_164_values3_0 = const()[name = string("concat_164_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_164_axis_0 = const()[name = string("concat_164_axis_0"), val = int32(0)];
            bool concat_164_interleave_0 = const()[name = string("concat_164_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_164 = concat(axis = concat_164_axis_0, interleave = concat_164_interleave_0, values = (expand_dims_245, concat_164_values1_0, cache_position_end, concat_164_values3_0))[name = string("concat_164")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_21_stride_0 = const()[name = string("key_cache_internal_tensor_assign_21_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_21_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_21_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_21_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_21_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_21_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_21_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_207_cast_fp16 = transpose(perm = key_states_207_perm_0, x = key_states_205_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_21_cast_fp16 = slice_update(begin = concat_163, begin_mask = key_cache_internal_tensor_assign_21_begin_mask_0, end = concat_164, end_mask = key_cache_internal_tensor_assign_21_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_21_squeeze_mask_0, stride = key_cache_internal_tensor_assign_21_stride_0, update = key_states_207_cast_fp16, x = coreml_update_state_94)[name = string("key_cache_internal_tensor_assign_21_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_21_cast_fp16, input = key_cache)[name = string("coreml_update_state_96_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_96 = read_state(input = key_cache)[name = string("coreml_update_state_96")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_21_stride_0 = const()[name = string("value_cache_internal_tensor_assign_21_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_21_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_21_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_21_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_21_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_21_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_21_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_123_cast_fp16 = transpose(perm = value_states_123_perm_0, x = var_7414_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_21_cast_fp16 = slice_update(begin = concat_163, begin_mask = value_cache_internal_tensor_assign_21_begin_mask_0, end = concat_164, end_mask = value_cache_internal_tensor_assign_21_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_21_squeeze_mask_0, stride = value_cache_internal_tensor_assign_21_stride_0, update = value_states_123_cast_fp16, x = coreml_update_state_95)[name = string("value_cache_internal_tensor_assign_21_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_21_cast_fp16, input = value_cache)[name = string("coreml_update_state_97_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_97 = read_state(input = value_cache)[name = string("coreml_update_state_97")];
            tensor<int32, [4]> var_7508_begin_0 = const()[name = string("op_7508_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_7508_end_0 = const()[name = string("op_7508_end_0"), val = tensor<int32, [4]>([21, 2, 2048, 128])];
            tensor<bool, [4]> var_7508_end_mask_0 = const()[name = string("op_7508_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_7508_cast_fp16 = slice_by_index(begin = var_7508_begin_0, end = var_7508_end_0, end_mask = var_7508_end_mask_0, x = coreml_update_state_96)[name = string("op_7508_cast_fp16")];
            tensor<int32, [2]> tile_40 = const()[name = string("tile_40"), val = tensor<int32, [2]>([1, 1])];
            int32 var_7511_axis_0 = const()[name = string("op_7511_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_7511_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_7511_cast_fp16_1 = split(axis = var_7511_axis_0, split_sizes = tile_40, x = var_7508_cast_fp16)[name = string("op_7511_cast_fp16")];
            tensor<int32, [4]> var_7518_begin_0 = const()[name = string("op_7518_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_7518_end_0 = const()[name = string("op_7518_end_0"), val = tensor<int32, [4]>([21, 2, 2048, 128])];
            tensor<bool, [4]> var_7518_end_mask_0 = const()[name = string("op_7518_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_7518_cast_fp16 = slice_by_index(begin = var_7518_begin_0, end = var_7518_end_0, end_mask = var_7518_end_mask_0, x = coreml_update_state_97)[name = string("op_7518_cast_fp16")];
            tensor<int32, [2]> tile_41 = const()[name = string("tile_41"), val = tensor<int32, [2]>([1, 1])];
            int32 var_7521_axis_0 = const()[name = string("op_7521_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_7521_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_7521_cast_fp16_1 = split(axis = var_7521_axis_0, split_sizes = tile_41, x = var_7518_cast_fp16)[name = string("op_7521_cast_fp16")];
            tensor<int32, [2]> var_7524_split_sizes_0 = const()[name = string("op_7524_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_7524_axis_0 = const()[name = string("op_7524_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_7524_0, tensor<fp16, [1, 8, 128, 8]> var_7524_1 = split(axis = var_7524_axis_0, split_sizes = var_7524_split_sizes_0, x = query_states_123_cast_fp16)[name = string("op_7524")];
            bool attn_weights_321_transpose_x_0 = const()[name = string("attn_weights_321_transpose_x_0"), val = bool(false)];
            bool attn_weights_321_transpose_y_0 = const()[name = string("attn_weights_321_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_321_cast_fp16 = matmul(transpose_x = attn_weights_321_transpose_x_0, transpose_y = attn_weights_321_transpose_y_0, x = var_7511_cast_fp16_0, y = var_7524_0)[name = string("attn_weights_321_cast_fp16")];
            fp16 var_7527_to_fp16 = const()[name = string("op_7527_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_323_cast_fp16 = mul(x = attn_weights_321_cast_fp16, y = var_7527_to_fp16)[name = string("attn_weights_323_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_325_cast_fp16 = add(x = attn_weights_323_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_325_cast_fp16")];
            int32 var_7531 = const()[name = string("op_7531"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_327_cast_fp16 = softmax(axis = var_7531, x = attn_weights_325_cast_fp16)[name = string("attn_weights_327_cast_fp16")];
            bool var_7537_transpose_x_1 = const()[name = string("op_7537_transpose_x_1"), val = bool(true)];
            bool var_7537_transpose_y_1 = const()[name = string("op_7537_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_7537_cast_fp16 = matmul(transpose_x = var_7537_transpose_x_1, transpose_y = var_7537_transpose_y_1, x = attn_weights_327_cast_fp16, y = var_7521_cast_fp16_0)[name = string("op_7537_cast_fp16")];
            bool attn_weights_329_transpose_x_0 = const()[name = string("attn_weights_329_transpose_x_0"), val = bool(false)];
            bool attn_weights_329_transpose_y_0 = const()[name = string("attn_weights_329_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_329_cast_fp16 = matmul(transpose_x = attn_weights_329_transpose_x_0, transpose_y = attn_weights_329_transpose_y_0, x = var_7511_cast_fp16_1, y = var_7524_1)[name = string("attn_weights_329_cast_fp16")];
            fp16 var_7539_to_fp16 = const()[name = string("op_7539_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_331_cast_fp16 = mul(x = attn_weights_329_cast_fp16, y = var_7539_to_fp16)[name = string("attn_weights_331_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_333_cast_fp16 = add(x = attn_weights_331_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_333_cast_fp16")];
            int32 var_7543 = const()[name = string("op_7543"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_335_cast_fp16 = softmax(axis = var_7543, x = attn_weights_333_cast_fp16)[name = string("attn_weights_335_cast_fp16")];
            bool attn_output_161_transpose_x_1 = const()[name = string("attn_output_161_transpose_x_1"), val = bool(true)];
            bool attn_output_161_transpose_y_1 = const()[name = string("attn_output_161_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_161_cast_fp16 = matmul(transpose_x = attn_output_161_transpose_x_1, transpose_y = attn_output_161_transpose_y_1, x = attn_weights_335_cast_fp16, y = var_7521_cast_fp16_1)[name = string("attn_output_161_cast_fp16")];
            int32 var_7551 = const()[name = string("op_7551"), val = int32(1)];
            bool attn_output_163_interleave_0 = const()[name = string("attn_output_163_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_163_cast_fp16 = concat(axis = var_7551, interleave = attn_output_163_interleave_0, values = (var_7537_cast_fp16, attn_output_161_cast_fp16))[name = string("attn_output_163_cast_fp16")];
            tensor<int32, [4]> var_7555_perm_0 = const()[name = string("op_7555_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_7560 = const()[name = string("op_7560"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_7555_cast_fp16 = transpose(perm = var_7555_perm_0, x = attn_output_163_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_167_cast_fp16 = reshape(shape = var_7560, x = var_7555_cast_fp16)[name = string("attn_output_167_cast_fp16")];
            string hidden_states_203_pad_type_0 = const()[name = string("hidden_states_203_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_203_strides_0 = const()[name = string("hidden_states_203_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_203_pad_0 = const()[name = string("hidden_states_203_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_203_dilations_0 = const()[name = string("hidden_states_203_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_203_groups_0 = const()[name = string("hidden_states_203_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_203_cast_fp16 = conv(dilations = hidden_states_203_dilations_0, groups = hidden_states_203_groups_0, pad = hidden_states_203_pad_0, pad_type = hidden_states_203_pad_type_0, strides = hidden_states_203_strides_0, weight = layers_20_self_attn_o_proj_weight_cast_fp16, x = attn_output_167_cast_fp16)[name = string("hidden_states_203_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_205_cast_fp16 = add(x = hidden_states_199_cast_fp16, y = hidden_states_203_cast_fp16)[name = string("hidden_states_205_cast_fp16")];
            int32 var_7586 = const()[name = string("op_7586"), val = int32(1)];
            fp16 const_253_promoted_to_fp16 = const()[name = string("const_253_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_7588_cast_fp16 = mul(x = hidden_states_205_cast_fp16, y = const_253_promoted_to_fp16)[name = string("op_7588_cast_fp16")];
            bool doubled_165_interleave_0 = const()[name = string("doubled_165_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_165_cast_fp16 = concat(axis = var_7586, interleave = doubled_165_interleave_0, values = (hidden_states_205_cast_fp16, var_7588_cast_fp16))[name = string("doubled_165_cast_fp16")];
            tensor<int32, [1]> out_83_axes_0 = const()[name = string("out_83_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_83_gamma_0_to_fp16 = const()[name = string("out_83_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1494199680)))];
            fp16 var_7598_to_fp16 = const()[name = string("op_7598_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_83_cast_fp16 = layer_norm(axes = out_83_axes_0, epsilon = var_7598_to_fp16, gamma = out_83_gamma_0_to_fp16, x = doubled_165_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<int32, [2]> var_7609_split_sizes_0 = const()[name = string("op_7609_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_7609_axis_0 = const()[name = string("op_7609_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_7609_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_7609_cast_fp16_1 = split(axis = var_7609_axis_0, split_sizes = var_7609_split_sizes_0, x = out_83_cast_fp16)[name = string("op_7609_cast_fp16")];
            string input_41_pad_type_0 = const()[name = string("input_41_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_41_strides_0 = const()[name = string("input_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_41_pad_0 = const()[name = string("input_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_41_dilations_0 = const()[name = string("input_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_41_groups_0 = const()[name = string("input_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_41_cast_fp16 = conv(dilations = input_41_dilations_0, groups = input_41_groups_0, pad = input_41_pad_0, pad_type = input_41_pad_type_0, strides = input_41_strides_0, weight = layers_20_mlp_gate_proj_weight_cast_fp16, x = var_7609_cast_fp16_0)[name = string("input_41_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_7626_cast_fp16 = silu(x = input_41_cast_fp16)[name = string("op_7626_cast_fp16")];
            string var_7632_pad_type_0 = const()[name = string("op_7632_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_7632_strides_0 = const()[name = string("op_7632_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7632_pad_0 = const()[name = string("op_7632_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7632_dilations_0 = const()[name = string("op_7632_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_7632_groups_0 = const()[name = string("op_7632_groups_0"), val = int32(1)];
            tensor<fp16, [6144, 2048, 1, 1]> layers_20_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_20_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1494207936)))];
            tensor<fp16, [1, 6144, 1, 8]> var_7632_cast_fp16 = conv(dilations = var_7632_dilations_0, groups = var_7632_groups_0, pad = var_7632_pad_0, pad_type = var_7632_pad_type_0, strides = var_7632_strides_0, weight = layers_20_mlp_up_proj_weight_to_fp16, x = var_7609_cast_fp16_0)[name = string("op_7632_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_209_cast_fp16 = mul(x = var_7626_cast_fp16, y = var_7632_cast_fp16)[name = string("x_209_cast_fp16")];
            string hidden_states_207_pad_type_0 = const()[name = string("hidden_states_207_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_207_strides_0 = const()[name = string("hidden_states_207_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_207_pad_0 = const()[name = string("hidden_states_207_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_207_dilations_0 = const()[name = string("hidden_states_207_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_207_groups_0 = const()[name = string("hidden_states_207_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_207_cast_fp16 = conv(dilations = hidden_states_207_dilations_0, groups = hidden_states_207_groups_0, pad = hidden_states_207_pad_0, pad_type = hidden_states_207_pad_type_0, strides = hidden_states_207_strides_0, weight = layers_20_mlp_down_proj_weight_cast_fp16, x = x_209_cast_fp16)[name = string("hidden_states_207_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_209_cast_fp16 = add(x = hidden_states_205_cast_fp16, y = hidden_states_207_cast_fp16)[name = string("hidden_states_209_cast_fp16")];
            int32 var_7648 = const()[name = string("op_7648"), val = int32(1)];
            fp16 const_255_promoted_to_fp16 = const()[name = string("const_255_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_7650_cast_fp16 = mul(x = hidden_states_209_cast_fp16, y = const_255_promoted_to_fp16)[name = string("op_7650_cast_fp16")];
            bool doubled_169_interleave_0 = const()[name = string("doubled_169_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_169_cast_fp16 = concat(axis = var_7648, interleave = doubled_169_interleave_0, values = (hidden_states_209_cast_fp16, var_7650_cast_fp16))[name = string("doubled_169_cast_fp16")];
            tensor<int32, [1]> out_85_axes_0 = const()[name = string("out_85_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_85_gamma_0_to_fp16 = const()[name = string("out_85_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1519373824)))];
            fp16 var_7660_to_fp16 = const()[name = string("op_7660_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_7660_to_fp16, gamma = out_85_gamma_0_to_fp16, x = doubled_169_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<int32, [2]> var_7671_split_sizes_0 = const()[name = string("op_7671_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_7671_axis_0 = const()[name = string("op_7671_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_7671_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_7671_cast_fp16_1 = split(axis = var_7671_axis_0, split_sizes = var_7671_split_sizes_0, x = out_85_cast_fp16)[name = string("op_7671_cast_fp16")];
            string query_states_127_pad_type_0 = const()[name = string("query_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_127_strides_0 = const()[name = string("query_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_127_pad_0 = const()[name = string("query_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_127_dilations_0 = const()[name = string("query_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_127_groups_0 = const()[name = string("query_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_127_cast_fp16 = conv(dilations = query_states_127_dilations_0, groups = query_states_127_groups_0, pad = query_states_127_pad_0, pad_type = query_states_127_pad_type_0, strides = query_states_127_strides_0, weight = layers_21_self_attn_q_proj_weight_cast_fp16, x = var_7671_cast_fp16_0)[name = string("query_states_127_cast_fp16")];
            string key_states_211_pad_type_0 = const()[name = string("key_states_211_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_211_strides_0 = const()[name = string("key_states_211_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_211_pad_0 = const()[name = string("key_states_211_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_211_dilations_0 = const()[name = string("key_states_211_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_211_groups_0 = const()[name = string("key_states_211_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_211_cast_fp16 = conv(dilations = key_states_211_dilations_0, groups = key_states_211_groups_0, pad = key_states_211_pad_0, pad_type = key_states_211_pad_type_0, strides = key_states_211_strides_0, weight = layers_21_self_attn_k_proj_weight_cast_fp16, x = var_7671_cast_fp16_0)[name = string("key_states_211_cast_fp16")];
            string value_states_127_pad_type_0 = const()[name = string("value_states_127_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_127_strides_0 = const()[name = string("value_states_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_127_pad_0 = const()[name = string("value_states_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_127_dilations_0 = const()[name = string("value_states_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_127_groups_0 = const()[name = string("value_states_127_groups_0"), val = int32(1)];
            tensor<fp16, [256, 2048, 1, 1]> layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1519382080)))];
            tensor<fp16, [1, 256, 1, 8]> value_states_127_cast_fp16 = conv(dilations = value_states_127_dilations_0, groups = value_states_127_groups_0, pad = value_states_127_pad_0, pad_type = value_states_127_pad_type_0, strides = value_states_127_strides_0, weight = layers_21_self_attn_v_proj_weight_to_fp16, x = var_7671_cast_fp16_0)[name = string("value_states_127_cast_fp16")];
            tensor<int32, [4]> var_7723 = const()[name = string("op_7723"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_211_cast_fp16 = reshape(shape = var_7723, x = query_states_127_cast_fp16)[name = string("x_211_cast_fp16")];
            tensor<int32, [4]> var_7727 = const()[name = string("op_7727"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_7728_cast_fp16 = reshape(shape = var_7727, x = key_states_211_cast_fp16)[name = string("op_7728_cast_fp16")];
            tensor<int32, [4]> var_7734 = const()[name = string("op_7734"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_7735_cast_fp16 = reshape(shape = var_7734, x = value_states_127_cast_fp16)[name = string("op_7735_cast_fp16")];
            tensor<int32, [4]> value_states_129_perm_0 = const()[name = string("value_states_129_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_7739_cast_fp16 = mul(x = x_211_cast_fp16, y = var_859_cast_fp16)[name = string("op_7739_cast_fp16")];
            tensor<int32, [2]> var_7740_split_sizes_0 = const()[name = string("op_7740_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_7740_axis_0 = const()[name = string("op_7740_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_7740_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_7740_cast_fp16_1 = split(axis = var_7740_axis_0, split_sizes = var_7740_split_sizes_0, x = x_211_cast_fp16)[name = string("op_7740_cast_fp16")];
            fp16 const_259_promoted_to_fp16 = const()[name = string("const_259_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_7742_cast_fp16 = mul(x = var_7740_cast_fp16_1, y = const_259_promoted_to_fp16)[name = string("op_7742_cast_fp16")];
            int32 var_7744 = const()[name = string("op_7744"), val = int32(-2)];
            bool var_7745_interleave_0 = const()[name = string("op_7745_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_7745_cast_fp16 = concat(axis = var_7744, interleave = var_7745_interleave_0, values = (var_7742_cast_fp16, var_7740_cast_fp16_0))[name = string("op_7745_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_7746_cast_fp16 = mul(x = var_7745_cast_fp16, y = var_866_cast_fp16)[name = string("op_7746_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_129_cast_fp16 = add(x = var_7739_cast_fp16, y = var_7746_cast_fp16)[name = string("query_states_129_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_7752_cast_fp16 = mul(x = var_7728_cast_fp16, y = var_859_cast_fp16)[name = string("op_7752_cast_fp16")];
            tensor<int32, [2]> var_7753_split_sizes_0 = const()[name = string("op_7753_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_7753_axis_0 = const()[name = string("op_7753_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_7753_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_7753_cast_fp16_1 = split(axis = var_7753_axis_0, split_sizes = var_7753_split_sizes_0, x = var_7728_cast_fp16)[name = string("op_7753_cast_fp16")];
            fp16 const_260_promoted_to_fp16 = const()[name = string("const_260_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_7755_cast_fp16 = mul(x = var_7753_cast_fp16_1, y = const_260_promoted_to_fp16)[name = string("op_7755_cast_fp16")];
            int32 var_7757 = const()[name = string("op_7757"), val = int32(-2)];
            bool var_7758_interleave_0 = const()[name = string("op_7758_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_7758_cast_fp16 = concat(axis = var_7757, interleave = var_7758_interleave_0, values = (var_7755_cast_fp16, var_7753_cast_fp16_0))[name = string("op_7758_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_7759_cast_fp16 = mul(x = var_7758_cast_fp16, y = var_866_cast_fp16)[name = string("op_7759_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_215_cast_fp16 = add(x = var_7752_cast_fp16, y = var_7759_cast_fp16)[name = string("key_states_215_cast_fp16")];
            tensor<int32, [4]> key_states_217_perm_0 = const()[name = string("key_states_217_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_253 = const()[name = string("expand_dims_253"), val = tensor<int32, [1]>([21])];
            tensor<int32, [1]> expand_dims_254 = const()[name = string("expand_dims_254"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_256 = const()[name = string("expand_dims_256"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_257 = const()[name = string("expand_dims_257"), val = tensor<int32, [1]>([22])];
            int32 concat_171_axis_0 = const()[name = string("concat_171_axis_0"), val = int32(0)];
            bool concat_171_interleave_0 = const()[name = string("concat_171_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_171 = concat(axis = concat_171_axis_0, interleave = concat_171_interleave_0, values = (expand_dims_253, expand_dims_254, position_id, expand_dims_256))[name = string("concat_171")];
            tensor<int32, [1]> concat_172_values1_0 = const()[name = string("concat_172_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_172_values3_0 = const()[name = string("concat_172_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_172_axis_0 = const()[name = string("concat_172_axis_0"), val = int32(0)];
            bool concat_172_interleave_0 = const()[name = string("concat_172_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_172 = concat(axis = concat_172_axis_0, interleave = concat_172_interleave_0, values = (expand_dims_257, concat_172_values1_0, cache_position_end, concat_172_values3_0))[name = string("concat_172")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_22_stride_0 = const()[name = string("key_cache_internal_tensor_assign_22_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_22_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_22_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_22_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_22_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_22_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_22_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_217_cast_fp16 = transpose(perm = key_states_217_perm_0, x = key_states_215_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_22_cast_fp16 = slice_update(begin = concat_171, begin_mask = key_cache_internal_tensor_assign_22_begin_mask_0, end = concat_172, end_mask = key_cache_internal_tensor_assign_22_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_22_squeeze_mask_0, stride = key_cache_internal_tensor_assign_22_stride_0, update = key_states_217_cast_fp16, x = coreml_update_state_96)[name = string("key_cache_internal_tensor_assign_22_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_22_cast_fp16, input = key_cache)[name = string("coreml_update_state_98_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_98 = read_state(input = key_cache)[name = string("coreml_update_state_98")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_22_stride_0 = const()[name = string("value_cache_internal_tensor_assign_22_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_22_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_22_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_22_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_22_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_22_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_22_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_129_cast_fp16 = transpose(perm = value_states_129_perm_0, x = var_7735_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_22_cast_fp16 = slice_update(begin = concat_171, begin_mask = value_cache_internal_tensor_assign_22_begin_mask_0, end = concat_172, end_mask = value_cache_internal_tensor_assign_22_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_22_squeeze_mask_0, stride = value_cache_internal_tensor_assign_22_stride_0, update = value_states_129_cast_fp16, x = coreml_update_state_97)[name = string("value_cache_internal_tensor_assign_22_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_22_cast_fp16, input = value_cache)[name = string("coreml_update_state_99_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_99 = read_state(input = value_cache)[name = string("coreml_update_state_99")];
            tensor<int32, [4]> var_7829_begin_0 = const()[name = string("op_7829_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_7829_end_0 = const()[name = string("op_7829_end_0"), val = tensor<int32, [4]>([22, 2, 2048, 128])];
            tensor<bool, [4]> var_7829_end_mask_0 = const()[name = string("op_7829_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_7829_cast_fp16 = slice_by_index(begin = var_7829_begin_0, end = var_7829_end_0, end_mask = var_7829_end_mask_0, x = coreml_update_state_98)[name = string("op_7829_cast_fp16")];
            tensor<int32, [2]> tile_42 = const()[name = string("tile_42"), val = tensor<int32, [2]>([1, 1])];
            int32 var_7832_axis_0 = const()[name = string("op_7832_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_7832_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_7832_cast_fp16_1 = split(axis = var_7832_axis_0, split_sizes = tile_42, x = var_7829_cast_fp16)[name = string("op_7832_cast_fp16")];
            tensor<int32, [4]> var_7839_begin_0 = const()[name = string("op_7839_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_7839_end_0 = const()[name = string("op_7839_end_0"), val = tensor<int32, [4]>([22, 2, 2048, 128])];
            tensor<bool, [4]> var_7839_end_mask_0 = const()[name = string("op_7839_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_7839_cast_fp16 = slice_by_index(begin = var_7839_begin_0, end = var_7839_end_0, end_mask = var_7839_end_mask_0, x = coreml_update_state_99)[name = string("op_7839_cast_fp16")];
            tensor<int32, [2]> tile_43 = const()[name = string("tile_43"), val = tensor<int32, [2]>([1, 1])];
            int32 var_7842_axis_0 = const()[name = string("op_7842_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_7842_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_7842_cast_fp16_1 = split(axis = var_7842_axis_0, split_sizes = tile_43, x = var_7839_cast_fp16)[name = string("op_7842_cast_fp16")];
            tensor<int32, [2]> var_7845_split_sizes_0 = const()[name = string("op_7845_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_7845_axis_0 = const()[name = string("op_7845_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_7845_0, tensor<fp16, [1, 8, 128, 8]> var_7845_1 = split(axis = var_7845_axis_0, split_sizes = var_7845_split_sizes_0, x = query_states_129_cast_fp16)[name = string("op_7845")];
            bool attn_weights_337_transpose_x_0 = const()[name = string("attn_weights_337_transpose_x_0"), val = bool(false)];
            bool attn_weights_337_transpose_y_0 = const()[name = string("attn_weights_337_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_337_cast_fp16 = matmul(transpose_x = attn_weights_337_transpose_x_0, transpose_y = attn_weights_337_transpose_y_0, x = var_7832_cast_fp16_0, y = var_7845_0)[name = string("attn_weights_337_cast_fp16")];
            fp16 var_7848_to_fp16 = const()[name = string("op_7848_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_339_cast_fp16 = mul(x = attn_weights_337_cast_fp16, y = var_7848_to_fp16)[name = string("attn_weights_339_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_341_cast_fp16 = add(x = attn_weights_339_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_341_cast_fp16")];
            int32 var_7852 = const()[name = string("op_7852"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_343_cast_fp16 = softmax(axis = var_7852, x = attn_weights_341_cast_fp16)[name = string("attn_weights_343_cast_fp16")];
            bool var_7858_transpose_x_1 = const()[name = string("op_7858_transpose_x_1"), val = bool(true)];
            bool var_7858_transpose_y_1 = const()[name = string("op_7858_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_7858_cast_fp16 = matmul(transpose_x = var_7858_transpose_x_1, transpose_y = var_7858_transpose_y_1, x = attn_weights_343_cast_fp16, y = var_7842_cast_fp16_0)[name = string("op_7858_cast_fp16")];
            bool attn_weights_345_transpose_x_0 = const()[name = string("attn_weights_345_transpose_x_0"), val = bool(false)];
            bool attn_weights_345_transpose_y_0 = const()[name = string("attn_weights_345_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_345_cast_fp16 = matmul(transpose_x = attn_weights_345_transpose_x_0, transpose_y = attn_weights_345_transpose_y_0, x = var_7832_cast_fp16_1, y = var_7845_1)[name = string("attn_weights_345_cast_fp16")];
            fp16 var_7860_to_fp16 = const()[name = string("op_7860_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_347_cast_fp16 = mul(x = attn_weights_345_cast_fp16, y = var_7860_to_fp16)[name = string("attn_weights_347_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_349_cast_fp16 = add(x = attn_weights_347_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_349_cast_fp16")];
            int32 var_7864 = const()[name = string("op_7864"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_351_cast_fp16 = softmax(axis = var_7864, x = attn_weights_349_cast_fp16)[name = string("attn_weights_351_cast_fp16")];
            bool attn_output_169_transpose_x_1 = const()[name = string("attn_output_169_transpose_x_1"), val = bool(true)];
            bool attn_output_169_transpose_y_1 = const()[name = string("attn_output_169_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_169_cast_fp16 = matmul(transpose_x = attn_output_169_transpose_x_1, transpose_y = attn_output_169_transpose_y_1, x = attn_weights_351_cast_fp16, y = var_7842_cast_fp16_1)[name = string("attn_output_169_cast_fp16")];
            int32 var_7872 = const()[name = string("op_7872"), val = int32(1)];
            bool attn_output_171_interleave_0 = const()[name = string("attn_output_171_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_171_cast_fp16 = concat(axis = var_7872, interleave = attn_output_171_interleave_0, values = (var_7858_cast_fp16, attn_output_169_cast_fp16))[name = string("attn_output_171_cast_fp16")];
            tensor<int32, [4]> var_7876_perm_0 = const()[name = string("op_7876_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_7881 = const()[name = string("op_7881"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_7876_cast_fp16 = transpose(perm = var_7876_perm_0, x = attn_output_171_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_175_cast_fp16 = reshape(shape = var_7881, x = var_7876_cast_fp16)[name = string("attn_output_175_cast_fp16")];
            string hidden_states_213_pad_type_0 = const()[name = string("hidden_states_213_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_213_strides_0 = const()[name = string("hidden_states_213_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_213_pad_0 = const()[name = string("hidden_states_213_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_213_dilations_0 = const()[name = string("hidden_states_213_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_213_groups_0 = const()[name = string("hidden_states_213_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_213_cast_fp16 = conv(dilations = hidden_states_213_dilations_0, groups = hidden_states_213_groups_0, pad = hidden_states_213_pad_0, pad_type = hidden_states_213_pad_type_0, strides = hidden_states_213_strides_0, weight = layers_21_self_attn_o_proj_weight_cast_fp16, x = attn_output_175_cast_fp16)[name = string("hidden_states_213_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_215_cast_fp16 = add(x = hidden_states_209_cast_fp16, y = hidden_states_213_cast_fp16)[name = string("hidden_states_215_cast_fp16")];
            int32 var_7907 = const()[name = string("op_7907"), val = int32(1)];
            fp16 const_265_promoted_to_fp16 = const()[name = string("const_265_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_7909_cast_fp16 = mul(x = hidden_states_215_cast_fp16, y = const_265_promoted_to_fp16)[name = string("op_7909_cast_fp16")];
            bool doubled_173_interleave_0 = const()[name = string("doubled_173_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_173_cast_fp16 = concat(axis = var_7907, interleave = doubled_173_interleave_0, values = (hidden_states_215_cast_fp16, var_7909_cast_fp16))[name = string("doubled_173_cast_fp16")];
            tensor<int32, [1]> out_87_axes_0 = const()[name = string("out_87_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_87_gamma_0_to_fp16 = const()[name = string("out_87_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1520430720)))];
            fp16 var_7919_to_fp16 = const()[name = string("op_7919_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_87_cast_fp16 = layer_norm(axes = out_87_axes_0, epsilon = var_7919_to_fp16, gamma = out_87_gamma_0_to_fp16, x = doubled_173_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<int32, [2]> var_7930_split_sizes_0 = const()[name = string("op_7930_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_7930_axis_0 = const()[name = string("op_7930_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_7930_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_7930_cast_fp16_1 = split(axis = var_7930_axis_0, split_sizes = var_7930_split_sizes_0, x = out_87_cast_fp16)[name = string("op_7930_cast_fp16")];
            string input_43_pad_type_0 = const()[name = string("input_43_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_43_strides_0 = const()[name = string("input_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_43_pad_0 = const()[name = string("input_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_43_dilations_0 = const()[name = string("input_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_43_groups_0 = const()[name = string("input_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_43_cast_fp16 = conv(dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = layers_21_mlp_gate_proj_weight_cast_fp16, x = var_7930_cast_fp16_0)[name = string("input_43_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_7947_cast_fp16 = silu(x = input_43_cast_fp16)[name = string("op_7947_cast_fp16")];
            string var_7953_pad_type_0 = const()[name = string("op_7953_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_7953_strides_0 = const()[name = string("op_7953_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7953_pad_0 = const()[name = string("op_7953_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7953_dilations_0 = const()[name = string("op_7953_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_7953_groups_0 = const()[name = string("op_7953_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_7953_cast_fp16 = conv(dilations = var_7953_dilations_0, groups = var_7953_groups_0, pad = var_7953_pad_0, pad_type = var_7953_pad_type_0, strides = var_7953_strides_0, weight = layers_21_mlp_up_proj_weight_cast_fp16, x = var_7930_cast_fp16_0)[name = string("op_7953_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_219_cast_fp16 = mul(x = var_7947_cast_fp16, y = var_7953_cast_fp16)[name = string("x_219_cast_fp16")];
            string hidden_states_217_pad_type_0 = const()[name = string("hidden_states_217_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_217_strides_0 = const()[name = string("hidden_states_217_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_217_pad_0 = const()[name = string("hidden_states_217_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_217_dilations_0 = const()[name = string("hidden_states_217_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_217_groups_0 = const()[name = string("hidden_states_217_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_217_cast_fp16 = conv(dilations = hidden_states_217_dilations_0, groups = hidden_states_217_groups_0, pad = hidden_states_217_pad_0, pad_type = hidden_states_217_pad_type_0, strides = hidden_states_217_strides_0, weight = layers_21_mlp_down_proj_weight_cast_fp16, x = x_219_cast_fp16)[name = string("hidden_states_217_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_219_cast_fp16 = add(x = hidden_states_215_cast_fp16, y = hidden_states_217_cast_fp16)[name = string("hidden_states_219_cast_fp16")];
            int32 var_7969 = const()[name = string("op_7969"), val = int32(1)];
            fp16 const_267_promoted_to_fp16 = const()[name = string("const_267_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_7971_cast_fp16 = mul(x = hidden_states_219_cast_fp16, y = const_267_promoted_to_fp16)[name = string("op_7971_cast_fp16")];
            bool doubled_177_interleave_0 = const()[name = string("doubled_177_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_177_cast_fp16 = concat(axis = var_7969, interleave = doubled_177_interleave_0, values = (hidden_states_219_cast_fp16, var_7971_cast_fp16))[name = string("doubled_177_cast_fp16")];
            tensor<int32, [1]> out_89_axes_0 = const()[name = string("out_89_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_89_gamma_0_to_fp16 = const()[name = string("out_89_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1520438976)))];
            fp16 var_7981_to_fp16 = const()[name = string("op_7981_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_89_cast_fp16 = layer_norm(axes = out_89_axes_0, epsilon = var_7981_to_fp16, gamma = out_89_gamma_0_to_fp16, x = doubled_177_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<int32, [2]> var_7992_split_sizes_0 = const()[name = string("op_7992_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_7992_axis_0 = const()[name = string("op_7992_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_7992_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_7992_cast_fp16_1 = split(axis = var_7992_axis_0, split_sizes = var_7992_split_sizes_0, x = out_89_cast_fp16)[name = string("op_7992_cast_fp16")];
            string query_states_133_pad_type_0 = const()[name = string("query_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_133_strides_0 = const()[name = string("query_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_133_pad_0 = const()[name = string("query_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_133_dilations_0 = const()[name = string("query_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_133_groups_0 = const()[name = string("query_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_133_cast_fp16 = conv(dilations = query_states_133_dilations_0, groups = query_states_133_groups_0, pad = query_states_133_pad_0, pad_type = query_states_133_pad_type_0, strides = query_states_133_strides_0, weight = layers_22_self_attn_q_proj_weight_cast_fp16, x = var_7992_cast_fp16_0)[name = string("query_states_133_cast_fp16")];
            string key_states_221_pad_type_0 = const()[name = string("key_states_221_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_221_strides_0 = const()[name = string("key_states_221_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_221_pad_0 = const()[name = string("key_states_221_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_221_dilations_0 = const()[name = string("key_states_221_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_221_groups_0 = const()[name = string("key_states_221_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_221_cast_fp16 = conv(dilations = key_states_221_dilations_0, groups = key_states_221_groups_0, pad = key_states_221_pad_0, pad_type = key_states_221_pad_type_0, strides = key_states_221_strides_0, weight = layers_22_self_attn_k_proj_weight_cast_fp16, x = var_7992_cast_fp16_0)[name = string("key_states_221_cast_fp16")];
            string value_states_133_pad_type_0 = const()[name = string("value_states_133_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_133_strides_0 = const()[name = string("value_states_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_133_pad_0 = const()[name = string("value_states_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_133_dilations_0 = const()[name = string("value_states_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_133_groups_0 = const()[name = string("value_states_133_groups_0"), val = int32(1)];
            tensor<fp16, [256, 2048, 1, 1]> layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1520447232)))];
            tensor<fp16, [1, 256, 1, 8]> value_states_133_cast_fp16 = conv(dilations = value_states_133_dilations_0, groups = value_states_133_groups_0, pad = value_states_133_pad_0, pad_type = value_states_133_pad_type_0, strides = value_states_133_strides_0, weight = layers_22_self_attn_v_proj_weight_to_fp16, x = var_7992_cast_fp16_0)[name = string("value_states_133_cast_fp16")];
            tensor<int32, [4]> var_8044 = const()[name = string("op_8044"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_221_cast_fp16 = reshape(shape = var_8044, x = query_states_133_cast_fp16)[name = string("x_221_cast_fp16")];
            tensor<int32, [4]> var_8048 = const()[name = string("op_8048"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_8049_cast_fp16 = reshape(shape = var_8048, x = key_states_221_cast_fp16)[name = string("op_8049_cast_fp16")];
            tensor<int32, [4]> var_8055 = const()[name = string("op_8055"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_8056_cast_fp16 = reshape(shape = var_8055, x = value_states_133_cast_fp16)[name = string("op_8056_cast_fp16")];
            tensor<int32, [4]> value_states_135_perm_0 = const()[name = string("value_states_135_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_8060_cast_fp16 = mul(x = x_221_cast_fp16, y = var_859_cast_fp16)[name = string("op_8060_cast_fp16")];
            tensor<int32, [2]> var_8061_split_sizes_0 = const()[name = string("op_8061_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_8061_axis_0 = const()[name = string("op_8061_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_8061_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_8061_cast_fp16_1 = split(axis = var_8061_axis_0, split_sizes = var_8061_split_sizes_0, x = x_221_cast_fp16)[name = string("op_8061_cast_fp16")];
            fp16 const_271_promoted_to_fp16 = const()[name = string("const_271_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_8063_cast_fp16 = mul(x = var_8061_cast_fp16_1, y = const_271_promoted_to_fp16)[name = string("op_8063_cast_fp16")];
            int32 var_8065 = const()[name = string("op_8065"), val = int32(-2)];
            bool var_8066_interleave_0 = const()[name = string("op_8066_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_8066_cast_fp16 = concat(axis = var_8065, interleave = var_8066_interleave_0, values = (var_8063_cast_fp16, var_8061_cast_fp16_0))[name = string("op_8066_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_8067_cast_fp16 = mul(x = var_8066_cast_fp16, y = var_866_cast_fp16)[name = string("op_8067_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_135_cast_fp16 = add(x = var_8060_cast_fp16, y = var_8067_cast_fp16)[name = string("query_states_135_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_8073_cast_fp16 = mul(x = var_8049_cast_fp16, y = var_859_cast_fp16)[name = string("op_8073_cast_fp16")];
            tensor<int32, [2]> var_8074_split_sizes_0 = const()[name = string("op_8074_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_8074_axis_0 = const()[name = string("op_8074_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_8074_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_8074_cast_fp16_1 = split(axis = var_8074_axis_0, split_sizes = var_8074_split_sizes_0, x = var_8049_cast_fp16)[name = string("op_8074_cast_fp16")];
            fp16 const_272_promoted_to_fp16 = const()[name = string("const_272_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_8076_cast_fp16 = mul(x = var_8074_cast_fp16_1, y = const_272_promoted_to_fp16)[name = string("op_8076_cast_fp16")];
            int32 var_8078 = const()[name = string("op_8078"), val = int32(-2)];
            bool var_8079_interleave_0 = const()[name = string("op_8079_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_8079_cast_fp16 = concat(axis = var_8078, interleave = var_8079_interleave_0, values = (var_8076_cast_fp16, var_8074_cast_fp16_0))[name = string("op_8079_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_8080_cast_fp16 = mul(x = var_8079_cast_fp16, y = var_866_cast_fp16)[name = string("op_8080_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_225_cast_fp16 = add(x = var_8073_cast_fp16, y = var_8080_cast_fp16)[name = string("key_states_225_cast_fp16")];
            tensor<int32, [4]> key_states_227_perm_0 = const()[name = string("key_states_227_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_265 = const()[name = string("expand_dims_265"), val = tensor<int32, [1]>([22])];
            tensor<int32, [1]> expand_dims_266 = const()[name = string("expand_dims_266"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_268 = const()[name = string("expand_dims_268"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_269 = const()[name = string("expand_dims_269"), val = tensor<int32, [1]>([23])];
            int32 concat_179_axis_0 = const()[name = string("concat_179_axis_0"), val = int32(0)];
            bool concat_179_interleave_0 = const()[name = string("concat_179_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_179 = concat(axis = concat_179_axis_0, interleave = concat_179_interleave_0, values = (expand_dims_265, expand_dims_266, position_id, expand_dims_268))[name = string("concat_179")];
            tensor<int32, [1]> concat_180_values1_0 = const()[name = string("concat_180_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_180_values3_0 = const()[name = string("concat_180_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_180_axis_0 = const()[name = string("concat_180_axis_0"), val = int32(0)];
            bool concat_180_interleave_0 = const()[name = string("concat_180_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_180 = concat(axis = concat_180_axis_0, interleave = concat_180_interleave_0, values = (expand_dims_269, concat_180_values1_0, cache_position_end, concat_180_values3_0))[name = string("concat_180")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_23_stride_0 = const()[name = string("key_cache_internal_tensor_assign_23_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_23_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_23_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_23_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_23_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_23_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_23_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_227_cast_fp16 = transpose(perm = key_states_227_perm_0, x = key_states_225_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_23_cast_fp16 = slice_update(begin = concat_179, begin_mask = key_cache_internal_tensor_assign_23_begin_mask_0, end = concat_180, end_mask = key_cache_internal_tensor_assign_23_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_23_squeeze_mask_0, stride = key_cache_internal_tensor_assign_23_stride_0, update = key_states_227_cast_fp16, x = coreml_update_state_98)[name = string("key_cache_internal_tensor_assign_23_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_23_cast_fp16, input = key_cache)[name = string("coreml_update_state_100_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_100 = read_state(input = key_cache)[name = string("coreml_update_state_100")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_23_stride_0 = const()[name = string("value_cache_internal_tensor_assign_23_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_23_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_23_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_23_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_23_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_23_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_23_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_135_cast_fp16 = transpose(perm = value_states_135_perm_0, x = var_8056_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_23_cast_fp16 = slice_update(begin = concat_179, begin_mask = value_cache_internal_tensor_assign_23_begin_mask_0, end = concat_180, end_mask = value_cache_internal_tensor_assign_23_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_23_squeeze_mask_0, stride = value_cache_internal_tensor_assign_23_stride_0, update = value_states_135_cast_fp16, x = coreml_update_state_99)[name = string("value_cache_internal_tensor_assign_23_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_23_cast_fp16, input = value_cache)[name = string("coreml_update_state_101_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_101 = read_state(input = value_cache)[name = string("coreml_update_state_101")];
            tensor<int32, [4]> var_8150_begin_0 = const()[name = string("op_8150_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8150_end_0 = const()[name = string("op_8150_end_0"), val = tensor<int32, [4]>([23, 2, 2048, 128])];
            tensor<bool, [4]> var_8150_end_mask_0 = const()[name = string("op_8150_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_8150_cast_fp16 = slice_by_index(begin = var_8150_begin_0, end = var_8150_end_0, end_mask = var_8150_end_mask_0, x = coreml_update_state_100)[name = string("op_8150_cast_fp16")];
            tensor<int32, [2]> tile_44 = const()[name = string("tile_44"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8153_axis_0 = const()[name = string("op_8153_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_8153_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_8153_cast_fp16_1 = split(axis = var_8153_axis_0, split_sizes = tile_44, x = var_8150_cast_fp16)[name = string("op_8153_cast_fp16")];
            tensor<int32, [4]> var_8160_begin_0 = const()[name = string("op_8160_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8160_end_0 = const()[name = string("op_8160_end_0"), val = tensor<int32, [4]>([23, 2, 2048, 128])];
            tensor<bool, [4]> var_8160_end_mask_0 = const()[name = string("op_8160_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_8160_cast_fp16 = slice_by_index(begin = var_8160_begin_0, end = var_8160_end_0, end_mask = var_8160_end_mask_0, x = coreml_update_state_101)[name = string("op_8160_cast_fp16")];
            tensor<int32, [2]> tile_45 = const()[name = string("tile_45"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8163_axis_0 = const()[name = string("op_8163_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_8163_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_8163_cast_fp16_1 = split(axis = var_8163_axis_0, split_sizes = tile_45, x = var_8160_cast_fp16)[name = string("op_8163_cast_fp16")];
            tensor<int32, [2]> var_8166_split_sizes_0 = const()[name = string("op_8166_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_8166_axis_0 = const()[name = string("op_8166_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_8166_0, tensor<fp16, [1, 8, 128, 8]> var_8166_1 = split(axis = var_8166_axis_0, split_sizes = var_8166_split_sizes_0, x = query_states_135_cast_fp16)[name = string("op_8166")];
            bool attn_weights_353_transpose_x_0 = const()[name = string("attn_weights_353_transpose_x_0"), val = bool(false)];
            bool attn_weights_353_transpose_y_0 = const()[name = string("attn_weights_353_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_353_cast_fp16 = matmul(transpose_x = attn_weights_353_transpose_x_0, transpose_y = attn_weights_353_transpose_y_0, x = var_8153_cast_fp16_0, y = var_8166_0)[name = string("attn_weights_353_cast_fp16")];
            fp16 var_8169_to_fp16 = const()[name = string("op_8169_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_355_cast_fp16 = mul(x = attn_weights_353_cast_fp16, y = var_8169_to_fp16)[name = string("attn_weights_355_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_357_cast_fp16 = add(x = attn_weights_355_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_357_cast_fp16")];
            int32 var_8173 = const()[name = string("op_8173"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_359_cast_fp16 = softmax(axis = var_8173, x = attn_weights_357_cast_fp16)[name = string("attn_weights_359_cast_fp16")];
            bool var_8179_transpose_x_1 = const()[name = string("op_8179_transpose_x_1"), val = bool(true)];
            bool var_8179_transpose_y_1 = const()[name = string("op_8179_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_8179_cast_fp16 = matmul(transpose_x = var_8179_transpose_x_1, transpose_y = var_8179_transpose_y_1, x = attn_weights_359_cast_fp16, y = var_8163_cast_fp16_0)[name = string("op_8179_cast_fp16")];
            bool attn_weights_361_transpose_x_0 = const()[name = string("attn_weights_361_transpose_x_0"), val = bool(false)];
            bool attn_weights_361_transpose_y_0 = const()[name = string("attn_weights_361_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_361_cast_fp16 = matmul(transpose_x = attn_weights_361_transpose_x_0, transpose_y = attn_weights_361_transpose_y_0, x = var_8153_cast_fp16_1, y = var_8166_1)[name = string("attn_weights_361_cast_fp16")];
            fp16 var_8181_to_fp16 = const()[name = string("op_8181_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_363_cast_fp16 = mul(x = attn_weights_361_cast_fp16, y = var_8181_to_fp16)[name = string("attn_weights_363_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_365_cast_fp16 = add(x = attn_weights_363_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_365_cast_fp16")];
            int32 var_8185 = const()[name = string("op_8185"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_367_cast_fp16 = softmax(axis = var_8185, x = attn_weights_365_cast_fp16)[name = string("attn_weights_367_cast_fp16")];
            bool attn_output_177_transpose_x_1 = const()[name = string("attn_output_177_transpose_x_1"), val = bool(true)];
            bool attn_output_177_transpose_y_1 = const()[name = string("attn_output_177_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_177_cast_fp16 = matmul(transpose_x = attn_output_177_transpose_x_1, transpose_y = attn_output_177_transpose_y_1, x = attn_weights_367_cast_fp16, y = var_8163_cast_fp16_1)[name = string("attn_output_177_cast_fp16")];
            int32 var_8193 = const()[name = string("op_8193"), val = int32(1)];
            bool attn_output_179_interleave_0 = const()[name = string("attn_output_179_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_179_cast_fp16 = concat(axis = var_8193, interleave = attn_output_179_interleave_0, values = (var_8179_cast_fp16, attn_output_177_cast_fp16))[name = string("attn_output_179_cast_fp16")];
            tensor<int32, [4]> var_8197_perm_0 = const()[name = string("op_8197_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_8202 = const()[name = string("op_8202"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_8197_cast_fp16 = transpose(perm = var_8197_perm_0, x = attn_output_179_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_183_cast_fp16 = reshape(shape = var_8202, x = var_8197_cast_fp16)[name = string("attn_output_183_cast_fp16")];
            string hidden_states_223_pad_type_0 = const()[name = string("hidden_states_223_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_223_strides_0 = const()[name = string("hidden_states_223_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_223_pad_0 = const()[name = string("hidden_states_223_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_223_dilations_0 = const()[name = string("hidden_states_223_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_223_groups_0 = const()[name = string("hidden_states_223_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 2048, 1, 1]> layers_22_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_22_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1521495872)))];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_223_cast_fp16 = conv(dilations = hidden_states_223_dilations_0, groups = hidden_states_223_groups_0, pad = hidden_states_223_pad_0, pad_type = hidden_states_223_pad_type_0, strides = hidden_states_223_strides_0, weight = layers_22_self_attn_o_proj_weight_to_fp16, x = attn_output_183_cast_fp16)[name = string("hidden_states_223_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_225_cast_fp16 = add(x = hidden_states_219_cast_fp16, y = hidden_states_223_cast_fp16)[name = string("hidden_states_225_cast_fp16")];
            int32 var_8228 = const()[name = string("op_8228"), val = int32(1)];
            fp16 const_277_promoted_to_fp16 = const()[name = string("const_277_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_8230_cast_fp16 = mul(x = hidden_states_225_cast_fp16, y = const_277_promoted_to_fp16)[name = string("op_8230_cast_fp16")];
            bool doubled_181_interleave_0 = const()[name = string("doubled_181_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_181_cast_fp16 = concat(axis = var_8228, interleave = doubled_181_interleave_0, values = (hidden_states_225_cast_fp16, var_8230_cast_fp16))[name = string("doubled_181_cast_fp16")];
            tensor<int32, [1]> out_91_axes_0 = const()[name = string("out_91_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_91_gamma_0_to_fp16 = const()[name = string("out_91_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1529884544)))];
            fp16 var_8240_to_fp16 = const()[name = string("op_8240_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_8240_to_fp16, gamma = out_91_gamma_0_to_fp16, x = doubled_181_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<int32, [2]> var_8251_split_sizes_0 = const()[name = string("op_8251_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_8251_axis_0 = const()[name = string("op_8251_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_8251_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_8251_cast_fp16_1 = split(axis = var_8251_axis_0, split_sizes = var_8251_split_sizes_0, x = out_91_cast_fp16)[name = string("op_8251_cast_fp16")];
            string input_45_pad_type_0 = const()[name = string("input_45_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_45_strides_0 = const()[name = string("input_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_45_pad_0 = const()[name = string("input_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_45_dilations_0 = const()[name = string("input_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_45_groups_0 = const()[name = string("input_45_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_45_cast_fp16 = conv(dilations = input_45_dilations_0, groups = input_45_groups_0, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = input_45_strides_0, weight = layers_22_mlp_gate_proj_weight_cast_fp16, x = var_8251_cast_fp16_0)[name = string("input_45_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_8268_cast_fp16 = silu(x = input_45_cast_fp16)[name = string("op_8268_cast_fp16")];
            string var_8274_pad_type_0 = const()[name = string("op_8274_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_8274_strides_0 = const()[name = string("op_8274_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8274_pad_0 = const()[name = string("op_8274_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8274_dilations_0 = const()[name = string("op_8274_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8274_groups_0 = const()[name = string("op_8274_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_8274_cast_fp16 = conv(dilations = var_8274_dilations_0, groups = var_8274_groups_0, pad = var_8274_pad_0, pad_type = var_8274_pad_type_0, strides = var_8274_strides_0, weight = layers_22_mlp_up_proj_weight_cast_fp16, x = var_8251_cast_fp16_0)[name = string("op_8274_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_229_cast_fp16 = mul(x = var_8268_cast_fp16, y = var_8274_cast_fp16)[name = string("x_229_cast_fp16")];
            string hidden_states_227_pad_type_0 = const()[name = string("hidden_states_227_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_227_strides_0 = const()[name = string("hidden_states_227_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_227_pad_0 = const()[name = string("hidden_states_227_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_227_dilations_0 = const()[name = string("hidden_states_227_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_227_groups_0 = const()[name = string("hidden_states_227_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_227_cast_fp16 = conv(dilations = hidden_states_227_dilations_0, groups = hidden_states_227_groups_0, pad = hidden_states_227_pad_0, pad_type = hidden_states_227_pad_type_0, strides = hidden_states_227_strides_0, weight = layers_22_mlp_down_proj_weight_cast_fp16, x = x_229_cast_fp16)[name = string("hidden_states_227_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_229_cast_fp16 = add(x = hidden_states_225_cast_fp16, y = hidden_states_227_cast_fp16)[name = string("hidden_states_229_cast_fp16")];
            int32 var_8290 = const()[name = string("op_8290"), val = int32(1)];
            fp16 const_279_promoted_to_fp16 = const()[name = string("const_279_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_8292_cast_fp16 = mul(x = hidden_states_229_cast_fp16, y = const_279_promoted_to_fp16)[name = string("op_8292_cast_fp16")];
            bool doubled_185_interleave_0 = const()[name = string("doubled_185_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_185_cast_fp16 = concat(axis = var_8290, interleave = doubled_185_interleave_0, values = (hidden_states_229_cast_fp16, var_8292_cast_fp16))[name = string("doubled_185_cast_fp16")];
            tensor<int32, [1]> out_93_axes_0 = const()[name = string("out_93_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_93_gamma_0_to_fp16 = const()[name = string("out_93_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1529892800)))];
            fp16 var_8302_to_fp16 = const()[name = string("op_8302_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_93_cast_fp16 = layer_norm(axes = out_93_axes_0, epsilon = var_8302_to_fp16, gamma = out_93_gamma_0_to_fp16, x = doubled_185_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<int32, [2]> var_8313_split_sizes_0 = const()[name = string("op_8313_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_8313_axis_0 = const()[name = string("op_8313_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_8313_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_8313_cast_fp16_1 = split(axis = var_8313_axis_0, split_sizes = var_8313_split_sizes_0, x = out_93_cast_fp16)[name = string("op_8313_cast_fp16")];
            string query_states_139_pad_type_0 = const()[name = string("query_states_139_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_139_strides_0 = const()[name = string("query_states_139_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_139_pad_0 = const()[name = string("query_states_139_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_139_dilations_0 = const()[name = string("query_states_139_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_139_groups_0 = const()[name = string("query_states_139_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_139_cast_fp16 = conv(dilations = query_states_139_dilations_0, groups = query_states_139_groups_0, pad = query_states_139_pad_0, pad_type = query_states_139_pad_type_0, strides = query_states_139_strides_0, weight = layers_23_self_attn_q_proj_weight_cast_fp16, x = var_8313_cast_fp16_0)[name = string("query_states_139_cast_fp16")];
            string key_states_231_pad_type_0 = const()[name = string("key_states_231_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_231_strides_0 = const()[name = string("key_states_231_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_231_pad_0 = const()[name = string("key_states_231_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_231_dilations_0 = const()[name = string("key_states_231_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_231_groups_0 = const()[name = string("key_states_231_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_231_cast_fp16 = conv(dilations = key_states_231_dilations_0, groups = key_states_231_groups_0, pad = key_states_231_pad_0, pad_type = key_states_231_pad_type_0, strides = key_states_231_strides_0, weight = layers_23_self_attn_k_proj_weight_cast_fp16, x = var_8313_cast_fp16_0)[name = string("key_states_231_cast_fp16")];
            string value_states_139_pad_type_0 = const()[name = string("value_states_139_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_139_strides_0 = const()[name = string("value_states_139_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_139_pad_0 = const()[name = string("value_states_139_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_139_dilations_0 = const()[name = string("value_states_139_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_139_groups_0 = const()[name = string("value_states_139_groups_0"), val = int32(1)];
            tensor<fp16, [256, 2048, 1, 1]> layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1529901056)))];
            tensor<fp16, [1, 256, 1, 8]> value_states_139_cast_fp16 = conv(dilations = value_states_139_dilations_0, groups = value_states_139_groups_0, pad = value_states_139_pad_0, pad_type = value_states_139_pad_type_0, strides = value_states_139_strides_0, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = var_8313_cast_fp16_0)[name = string("value_states_139_cast_fp16")];
            tensor<int32, [4]> var_8365 = const()[name = string("op_8365"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_231_cast_fp16 = reshape(shape = var_8365, x = query_states_139_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<int32, [4]> var_8369 = const()[name = string("op_8369"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_8370_cast_fp16 = reshape(shape = var_8369, x = key_states_231_cast_fp16)[name = string("op_8370_cast_fp16")];
            tensor<int32, [4]> var_8376 = const()[name = string("op_8376"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_8377_cast_fp16 = reshape(shape = var_8376, x = value_states_139_cast_fp16)[name = string("op_8377_cast_fp16")];
            tensor<int32, [4]> value_states_141_perm_0 = const()[name = string("value_states_141_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_8381_cast_fp16 = mul(x = x_231_cast_fp16, y = var_859_cast_fp16)[name = string("op_8381_cast_fp16")];
            tensor<int32, [2]> var_8382_split_sizes_0 = const()[name = string("op_8382_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_8382_axis_0 = const()[name = string("op_8382_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_8382_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_8382_cast_fp16_1 = split(axis = var_8382_axis_0, split_sizes = var_8382_split_sizes_0, x = x_231_cast_fp16)[name = string("op_8382_cast_fp16")];
            fp16 const_283_promoted_to_fp16 = const()[name = string("const_283_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_8384_cast_fp16 = mul(x = var_8382_cast_fp16_1, y = const_283_promoted_to_fp16)[name = string("op_8384_cast_fp16")];
            int32 var_8386 = const()[name = string("op_8386"), val = int32(-2)];
            bool var_8387_interleave_0 = const()[name = string("op_8387_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_8387_cast_fp16 = concat(axis = var_8386, interleave = var_8387_interleave_0, values = (var_8384_cast_fp16, var_8382_cast_fp16_0))[name = string("op_8387_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_8388_cast_fp16 = mul(x = var_8387_cast_fp16, y = var_866_cast_fp16)[name = string("op_8388_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_141_cast_fp16 = add(x = var_8381_cast_fp16, y = var_8388_cast_fp16)[name = string("query_states_141_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_8394_cast_fp16 = mul(x = var_8370_cast_fp16, y = var_859_cast_fp16)[name = string("op_8394_cast_fp16")];
            tensor<int32, [2]> var_8395_split_sizes_0 = const()[name = string("op_8395_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_8395_axis_0 = const()[name = string("op_8395_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_8395_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_8395_cast_fp16_1 = split(axis = var_8395_axis_0, split_sizes = var_8395_split_sizes_0, x = var_8370_cast_fp16)[name = string("op_8395_cast_fp16")];
            fp16 const_284_promoted_to_fp16 = const()[name = string("const_284_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_8397_cast_fp16 = mul(x = var_8395_cast_fp16_1, y = const_284_promoted_to_fp16)[name = string("op_8397_cast_fp16")];
            int32 var_8399 = const()[name = string("op_8399"), val = int32(-2)];
            bool var_8400_interleave_0 = const()[name = string("op_8400_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_8400_cast_fp16 = concat(axis = var_8399, interleave = var_8400_interleave_0, values = (var_8397_cast_fp16, var_8395_cast_fp16_0))[name = string("op_8400_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_8401_cast_fp16 = mul(x = var_8400_cast_fp16, y = var_866_cast_fp16)[name = string("op_8401_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_235_cast_fp16 = add(x = var_8394_cast_fp16, y = var_8401_cast_fp16)[name = string("key_states_235_cast_fp16")];
            tensor<int32, [4]> key_states_237_perm_0 = const()[name = string("key_states_237_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_277 = const()[name = string("expand_dims_277"), val = tensor<int32, [1]>([23])];
            tensor<int32, [1]> expand_dims_278 = const()[name = string("expand_dims_278"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_280 = const()[name = string("expand_dims_280"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_281 = const()[name = string("expand_dims_281"), val = tensor<int32, [1]>([24])];
            int32 concat_187_axis_0 = const()[name = string("concat_187_axis_0"), val = int32(0)];
            bool concat_187_interleave_0 = const()[name = string("concat_187_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_187 = concat(axis = concat_187_axis_0, interleave = concat_187_interleave_0, values = (expand_dims_277, expand_dims_278, position_id, expand_dims_280))[name = string("concat_187")];
            tensor<int32, [1]> concat_188_values1_0 = const()[name = string("concat_188_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_188_values3_0 = const()[name = string("concat_188_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_188_axis_0 = const()[name = string("concat_188_axis_0"), val = int32(0)];
            bool concat_188_interleave_0 = const()[name = string("concat_188_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_188 = concat(axis = concat_188_axis_0, interleave = concat_188_interleave_0, values = (expand_dims_281, concat_188_values1_0, cache_position_end, concat_188_values3_0))[name = string("concat_188")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_24_stride_0 = const()[name = string("key_cache_internal_tensor_assign_24_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_24_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_24_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_24_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_24_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_24_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_24_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_237_cast_fp16 = transpose(perm = key_states_237_perm_0, x = key_states_235_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_24_cast_fp16 = slice_update(begin = concat_187, begin_mask = key_cache_internal_tensor_assign_24_begin_mask_0, end = concat_188, end_mask = key_cache_internal_tensor_assign_24_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_24_squeeze_mask_0, stride = key_cache_internal_tensor_assign_24_stride_0, update = key_states_237_cast_fp16, x = coreml_update_state_100)[name = string("key_cache_internal_tensor_assign_24_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_24_cast_fp16, input = key_cache)[name = string("coreml_update_state_102_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_102 = read_state(input = key_cache)[name = string("coreml_update_state_102")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_24_stride_0 = const()[name = string("value_cache_internal_tensor_assign_24_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_24_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_24_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_24_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_24_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_24_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_24_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_141_cast_fp16 = transpose(perm = value_states_141_perm_0, x = var_8377_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_24_cast_fp16 = slice_update(begin = concat_187, begin_mask = value_cache_internal_tensor_assign_24_begin_mask_0, end = concat_188, end_mask = value_cache_internal_tensor_assign_24_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_24_squeeze_mask_0, stride = value_cache_internal_tensor_assign_24_stride_0, update = value_states_141_cast_fp16, x = coreml_update_state_101)[name = string("value_cache_internal_tensor_assign_24_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_24_cast_fp16, input = value_cache)[name = string("coreml_update_state_103_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_103 = read_state(input = value_cache)[name = string("coreml_update_state_103")];
            tensor<int32, [4]> var_8471_begin_0 = const()[name = string("op_8471_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_8471_end_0 = const()[name = string("op_8471_end_0"), val = tensor<int32, [4]>([24, 2, 2048, 128])];
            tensor<bool, [4]> var_8471_end_mask_0 = const()[name = string("op_8471_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_8471_cast_fp16 = slice_by_index(begin = var_8471_begin_0, end = var_8471_end_0, end_mask = var_8471_end_mask_0, x = coreml_update_state_102)[name = string("op_8471_cast_fp16")];
            tensor<int32, [2]> tile_46 = const()[name = string("tile_46"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8474_axis_0 = const()[name = string("op_8474_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_8474_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_8474_cast_fp16_1 = split(axis = var_8474_axis_0, split_sizes = tile_46, x = var_8471_cast_fp16)[name = string("op_8474_cast_fp16")];
            tensor<int32, [4]> var_8481_begin_0 = const()[name = string("op_8481_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_8481_end_0 = const()[name = string("op_8481_end_0"), val = tensor<int32, [4]>([24, 2, 2048, 128])];
            tensor<bool, [4]> var_8481_end_mask_0 = const()[name = string("op_8481_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_8481_cast_fp16 = slice_by_index(begin = var_8481_begin_0, end = var_8481_end_0, end_mask = var_8481_end_mask_0, x = coreml_update_state_103)[name = string("op_8481_cast_fp16")];
            tensor<int32, [2]> tile_47 = const()[name = string("tile_47"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8484_axis_0 = const()[name = string("op_8484_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_8484_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_8484_cast_fp16_1 = split(axis = var_8484_axis_0, split_sizes = tile_47, x = var_8481_cast_fp16)[name = string("op_8484_cast_fp16")];
            tensor<int32, [2]> var_8487_split_sizes_0 = const()[name = string("op_8487_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_8487_axis_0 = const()[name = string("op_8487_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_8487_0, tensor<fp16, [1, 8, 128, 8]> var_8487_1 = split(axis = var_8487_axis_0, split_sizes = var_8487_split_sizes_0, x = query_states_141_cast_fp16)[name = string("op_8487")];
            bool attn_weights_369_transpose_x_0 = const()[name = string("attn_weights_369_transpose_x_0"), val = bool(false)];
            bool attn_weights_369_transpose_y_0 = const()[name = string("attn_weights_369_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_369_cast_fp16 = matmul(transpose_x = attn_weights_369_transpose_x_0, transpose_y = attn_weights_369_transpose_y_0, x = var_8474_cast_fp16_0, y = var_8487_0)[name = string("attn_weights_369_cast_fp16")];
            fp16 var_8490_to_fp16 = const()[name = string("op_8490_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_371_cast_fp16 = mul(x = attn_weights_369_cast_fp16, y = var_8490_to_fp16)[name = string("attn_weights_371_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_373_cast_fp16 = add(x = attn_weights_371_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_373_cast_fp16")];
            int32 var_8494 = const()[name = string("op_8494"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_375_cast_fp16 = softmax(axis = var_8494, x = attn_weights_373_cast_fp16)[name = string("attn_weights_375_cast_fp16")];
            bool var_8500_transpose_x_1 = const()[name = string("op_8500_transpose_x_1"), val = bool(true)];
            bool var_8500_transpose_y_1 = const()[name = string("op_8500_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_8500_cast_fp16 = matmul(transpose_x = var_8500_transpose_x_1, transpose_y = var_8500_transpose_y_1, x = attn_weights_375_cast_fp16, y = var_8484_cast_fp16_0)[name = string("op_8500_cast_fp16")];
            bool attn_weights_377_transpose_x_0 = const()[name = string("attn_weights_377_transpose_x_0"), val = bool(false)];
            bool attn_weights_377_transpose_y_0 = const()[name = string("attn_weights_377_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_377_cast_fp16 = matmul(transpose_x = attn_weights_377_transpose_x_0, transpose_y = attn_weights_377_transpose_y_0, x = var_8474_cast_fp16_1, y = var_8487_1)[name = string("attn_weights_377_cast_fp16")];
            fp16 var_8502_to_fp16 = const()[name = string("op_8502_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_379_cast_fp16 = mul(x = attn_weights_377_cast_fp16, y = var_8502_to_fp16)[name = string("attn_weights_379_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_381_cast_fp16 = add(x = attn_weights_379_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_381_cast_fp16")];
            int32 var_8506 = const()[name = string("op_8506"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_383_cast_fp16 = softmax(axis = var_8506, x = attn_weights_381_cast_fp16)[name = string("attn_weights_383_cast_fp16")];
            bool attn_output_185_transpose_x_1 = const()[name = string("attn_output_185_transpose_x_1"), val = bool(true)];
            bool attn_output_185_transpose_y_1 = const()[name = string("attn_output_185_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_185_cast_fp16 = matmul(transpose_x = attn_output_185_transpose_x_1, transpose_y = attn_output_185_transpose_y_1, x = attn_weights_383_cast_fp16, y = var_8484_cast_fp16_1)[name = string("attn_output_185_cast_fp16")];
            int32 var_8514 = const()[name = string("op_8514"), val = int32(1)];
            bool attn_output_187_interleave_0 = const()[name = string("attn_output_187_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_187_cast_fp16 = concat(axis = var_8514, interleave = attn_output_187_interleave_0, values = (var_8500_cast_fp16, attn_output_185_cast_fp16))[name = string("attn_output_187_cast_fp16")];
            tensor<int32, [4]> var_8518_perm_0 = const()[name = string("op_8518_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_8523 = const()[name = string("op_8523"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_8518_cast_fp16 = transpose(perm = var_8518_perm_0, x = attn_output_187_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_191_cast_fp16 = reshape(shape = var_8523, x = var_8518_cast_fp16)[name = string("attn_output_191_cast_fp16")];
            string hidden_states_233_pad_type_0 = const()[name = string("hidden_states_233_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_233_strides_0 = const()[name = string("hidden_states_233_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_233_pad_0 = const()[name = string("hidden_states_233_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_233_dilations_0 = const()[name = string("hidden_states_233_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_233_groups_0 = const()[name = string("hidden_states_233_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_233_cast_fp16 = conv(dilations = hidden_states_233_dilations_0, groups = hidden_states_233_groups_0, pad = hidden_states_233_pad_0, pad_type = hidden_states_233_pad_type_0, strides = hidden_states_233_strides_0, weight = layers_23_self_attn_o_proj_weight_cast_fp16, x = attn_output_191_cast_fp16)[name = string("hidden_states_233_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_235_cast_fp16 = add(x = hidden_states_229_cast_fp16, y = hidden_states_233_cast_fp16)[name = string("hidden_states_235_cast_fp16")];
            int32 var_8549 = const()[name = string("op_8549"), val = int32(1)];
            fp16 const_289_promoted_to_fp16 = const()[name = string("const_289_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_8551_cast_fp16 = mul(x = hidden_states_235_cast_fp16, y = const_289_promoted_to_fp16)[name = string("op_8551_cast_fp16")];
            bool doubled_189_interleave_0 = const()[name = string("doubled_189_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_189_cast_fp16 = concat(axis = var_8549, interleave = doubled_189_interleave_0, values = (hidden_states_235_cast_fp16, var_8551_cast_fp16))[name = string("doubled_189_cast_fp16")];
            tensor<int32, [1]> out_95_axes_0 = const()[name = string("out_95_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_95_gamma_0_to_fp16 = const()[name = string("out_95_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1530949696)))];
            fp16 var_8561_to_fp16 = const()[name = string("op_8561_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_95_cast_fp16 = layer_norm(axes = out_95_axes_0, epsilon = var_8561_to_fp16, gamma = out_95_gamma_0_to_fp16, x = doubled_189_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<int32, [2]> var_8572_split_sizes_0 = const()[name = string("op_8572_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_8572_axis_0 = const()[name = string("op_8572_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_8572_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_8572_cast_fp16_1 = split(axis = var_8572_axis_0, split_sizes = var_8572_split_sizes_0, x = out_95_cast_fp16)[name = string("op_8572_cast_fp16")];
            string input_47_pad_type_0 = const()[name = string("input_47_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_47_strides_0 = const()[name = string("input_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_47_pad_0 = const()[name = string("input_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_47_dilations_0 = const()[name = string("input_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_47_groups_0 = const()[name = string("input_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_47_cast_fp16 = conv(dilations = input_47_dilations_0, groups = input_47_groups_0, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = input_47_strides_0, weight = layers_23_mlp_gate_proj_weight_cast_fp16, x = var_8572_cast_fp16_0)[name = string("input_47_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_8589_cast_fp16 = silu(x = input_47_cast_fp16)[name = string("op_8589_cast_fp16")];
            string var_8595_pad_type_0 = const()[name = string("op_8595_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_8595_strides_0 = const()[name = string("op_8595_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8595_pad_0 = const()[name = string("op_8595_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8595_dilations_0 = const()[name = string("op_8595_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8595_groups_0 = const()[name = string("op_8595_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_8595_cast_fp16 = conv(dilations = var_8595_dilations_0, groups = var_8595_groups_0, pad = var_8595_pad_0, pad_type = var_8595_pad_type_0, strides = var_8595_strides_0, weight = layers_23_mlp_up_proj_weight_cast_fp16, x = var_8572_cast_fp16_0)[name = string("op_8595_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_239_cast_fp16 = mul(x = var_8589_cast_fp16, y = var_8595_cast_fp16)[name = string("x_239_cast_fp16")];
            string hidden_states_237_pad_type_0 = const()[name = string("hidden_states_237_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_237_strides_0 = const()[name = string("hidden_states_237_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_237_pad_0 = const()[name = string("hidden_states_237_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_237_dilations_0 = const()[name = string("hidden_states_237_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_237_groups_0 = const()[name = string("hidden_states_237_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_237_cast_fp16 = conv(dilations = hidden_states_237_dilations_0, groups = hidden_states_237_groups_0, pad = hidden_states_237_pad_0, pad_type = hidden_states_237_pad_type_0, strides = hidden_states_237_strides_0, weight = layers_23_mlp_down_proj_weight_cast_fp16, x = x_239_cast_fp16)[name = string("hidden_states_237_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_239_cast_fp16 = add(x = hidden_states_235_cast_fp16, y = hidden_states_237_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            int32 var_8611 = const()[name = string("op_8611"), val = int32(1)];
            fp16 const_291_promoted_to_fp16 = const()[name = string("const_291_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_8613_cast_fp16 = mul(x = hidden_states_239_cast_fp16, y = const_291_promoted_to_fp16)[name = string("op_8613_cast_fp16")];
            bool doubled_193_interleave_0 = const()[name = string("doubled_193_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_193_cast_fp16 = concat(axis = var_8611, interleave = doubled_193_interleave_0, values = (hidden_states_239_cast_fp16, var_8613_cast_fp16))[name = string("doubled_193_cast_fp16")];
            tensor<int32, [1]> out_97_axes_0 = const()[name = string("out_97_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_97_gamma_0_to_fp16 = const()[name = string("out_97_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1530957952)))];
            fp16 var_8623_to_fp16 = const()[name = string("op_8623_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_8623_to_fp16, gamma = out_97_gamma_0_to_fp16, x = doubled_193_cast_fp16)[name = string("out_97_cast_fp16")];
            tensor<int32, [2]> var_8634_split_sizes_0 = const()[name = string("op_8634_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_8634_axis_0 = const()[name = string("op_8634_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_8634_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_8634_cast_fp16_1 = split(axis = var_8634_axis_0, split_sizes = var_8634_split_sizes_0, x = out_97_cast_fp16)[name = string("op_8634_cast_fp16")];
            string query_states_145_pad_type_0 = const()[name = string("query_states_145_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_145_strides_0 = const()[name = string("query_states_145_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_145_pad_0 = const()[name = string("query_states_145_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_145_dilations_0 = const()[name = string("query_states_145_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_145_groups_0 = const()[name = string("query_states_145_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_145_cast_fp16 = conv(dilations = query_states_145_dilations_0, groups = query_states_145_groups_0, pad = query_states_145_pad_0, pad_type = query_states_145_pad_type_0, strides = query_states_145_strides_0, weight = layers_24_self_attn_q_proj_weight_cast_fp16, x = var_8634_cast_fp16_0)[name = string("query_states_145_cast_fp16")];
            string key_states_241_pad_type_0 = const()[name = string("key_states_241_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_241_strides_0 = const()[name = string("key_states_241_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_241_pad_0 = const()[name = string("key_states_241_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_241_dilations_0 = const()[name = string("key_states_241_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_241_groups_0 = const()[name = string("key_states_241_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_241_cast_fp16 = conv(dilations = key_states_241_dilations_0, groups = key_states_241_groups_0, pad = key_states_241_pad_0, pad_type = key_states_241_pad_type_0, strides = key_states_241_strides_0, weight = layers_24_self_attn_k_proj_weight_cast_fp16, x = var_8634_cast_fp16_0)[name = string("key_states_241_cast_fp16")];
            string value_states_145_pad_type_0 = const()[name = string("value_states_145_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_145_strides_0 = const()[name = string("value_states_145_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_145_pad_0 = const()[name = string("value_states_145_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_145_dilations_0 = const()[name = string("value_states_145_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_145_groups_0 = const()[name = string("value_states_145_groups_0"), val = int32(1)];
            tensor<fp16, [256, 2048, 1, 1]> layers_24_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_24_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1530966208)))];
            tensor<fp16, [1, 256, 1, 8]> value_states_145_cast_fp16 = conv(dilations = value_states_145_dilations_0, groups = value_states_145_groups_0, pad = value_states_145_pad_0, pad_type = value_states_145_pad_type_0, strides = value_states_145_strides_0, weight = layers_24_self_attn_v_proj_weight_to_fp16, x = var_8634_cast_fp16_0)[name = string("value_states_145_cast_fp16")];
            tensor<int32, [4]> var_8686 = const()[name = string("op_8686"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_241_cast_fp16 = reshape(shape = var_8686, x = query_states_145_cast_fp16)[name = string("x_241_cast_fp16")];
            tensor<int32, [4]> var_8690 = const()[name = string("op_8690"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_8691_cast_fp16 = reshape(shape = var_8690, x = key_states_241_cast_fp16)[name = string("op_8691_cast_fp16")];
            tensor<int32, [4]> var_8697 = const()[name = string("op_8697"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_8698_cast_fp16 = reshape(shape = var_8697, x = value_states_145_cast_fp16)[name = string("op_8698_cast_fp16")];
            tensor<int32, [4]> value_states_147_perm_0 = const()[name = string("value_states_147_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_8702_cast_fp16 = mul(x = x_241_cast_fp16, y = var_859_cast_fp16)[name = string("op_8702_cast_fp16")];
            tensor<int32, [2]> var_8703_split_sizes_0 = const()[name = string("op_8703_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_8703_axis_0 = const()[name = string("op_8703_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_8703_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_8703_cast_fp16_1 = split(axis = var_8703_axis_0, split_sizes = var_8703_split_sizes_0, x = x_241_cast_fp16)[name = string("op_8703_cast_fp16")];
            fp16 const_295_promoted_to_fp16 = const()[name = string("const_295_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_8705_cast_fp16 = mul(x = var_8703_cast_fp16_1, y = const_295_promoted_to_fp16)[name = string("op_8705_cast_fp16")];
            int32 var_8707 = const()[name = string("op_8707"), val = int32(-2)];
            bool var_8708_interleave_0 = const()[name = string("op_8708_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_8708_cast_fp16 = concat(axis = var_8707, interleave = var_8708_interleave_0, values = (var_8705_cast_fp16, var_8703_cast_fp16_0))[name = string("op_8708_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_8709_cast_fp16 = mul(x = var_8708_cast_fp16, y = var_866_cast_fp16)[name = string("op_8709_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_147_cast_fp16 = add(x = var_8702_cast_fp16, y = var_8709_cast_fp16)[name = string("query_states_147_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_8715_cast_fp16 = mul(x = var_8691_cast_fp16, y = var_859_cast_fp16)[name = string("op_8715_cast_fp16")];
            tensor<int32, [2]> var_8716_split_sizes_0 = const()[name = string("op_8716_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_8716_axis_0 = const()[name = string("op_8716_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_8716_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_8716_cast_fp16_1 = split(axis = var_8716_axis_0, split_sizes = var_8716_split_sizes_0, x = var_8691_cast_fp16)[name = string("op_8716_cast_fp16")];
            fp16 const_296_promoted_to_fp16 = const()[name = string("const_296_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_8718_cast_fp16 = mul(x = var_8716_cast_fp16_1, y = const_296_promoted_to_fp16)[name = string("op_8718_cast_fp16")];
            int32 var_8720 = const()[name = string("op_8720"), val = int32(-2)];
            bool var_8721_interleave_0 = const()[name = string("op_8721_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_8721_cast_fp16 = concat(axis = var_8720, interleave = var_8721_interleave_0, values = (var_8718_cast_fp16, var_8716_cast_fp16_0))[name = string("op_8721_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_8722_cast_fp16 = mul(x = var_8721_cast_fp16, y = var_866_cast_fp16)[name = string("op_8722_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_245_cast_fp16 = add(x = var_8715_cast_fp16, y = var_8722_cast_fp16)[name = string("key_states_245_cast_fp16")];
            tensor<int32, [4]> key_states_247_perm_0 = const()[name = string("key_states_247_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_289 = const()[name = string("expand_dims_289"), val = tensor<int32, [1]>([24])];
            tensor<int32, [1]> expand_dims_290 = const()[name = string("expand_dims_290"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_292 = const()[name = string("expand_dims_292"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_293 = const()[name = string("expand_dims_293"), val = tensor<int32, [1]>([25])];
            int32 concat_195_axis_0 = const()[name = string("concat_195_axis_0"), val = int32(0)];
            bool concat_195_interleave_0 = const()[name = string("concat_195_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_195 = concat(axis = concat_195_axis_0, interleave = concat_195_interleave_0, values = (expand_dims_289, expand_dims_290, position_id, expand_dims_292))[name = string("concat_195")];
            tensor<int32, [1]> concat_196_values1_0 = const()[name = string("concat_196_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_196_values3_0 = const()[name = string("concat_196_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_196_axis_0 = const()[name = string("concat_196_axis_0"), val = int32(0)];
            bool concat_196_interleave_0 = const()[name = string("concat_196_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_196 = concat(axis = concat_196_axis_0, interleave = concat_196_interleave_0, values = (expand_dims_293, concat_196_values1_0, cache_position_end, concat_196_values3_0))[name = string("concat_196")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_25_stride_0 = const()[name = string("key_cache_internal_tensor_assign_25_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_25_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_25_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_25_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_25_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_25_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_25_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_247_cast_fp16 = transpose(perm = key_states_247_perm_0, x = key_states_245_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_25_cast_fp16 = slice_update(begin = concat_195, begin_mask = key_cache_internal_tensor_assign_25_begin_mask_0, end = concat_196, end_mask = key_cache_internal_tensor_assign_25_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_25_squeeze_mask_0, stride = key_cache_internal_tensor_assign_25_stride_0, update = key_states_247_cast_fp16, x = coreml_update_state_102)[name = string("key_cache_internal_tensor_assign_25_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_25_cast_fp16, input = key_cache)[name = string("coreml_update_state_104_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_104 = read_state(input = key_cache)[name = string("coreml_update_state_104")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_25_stride_0 = const()[name = string("value_cache_internal_tensor_assign_25_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_25_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_25_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_25_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_25_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_25_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_25_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_147_cast_fp16 = transpose(perm = value_states_147_perm_0, x = var_8698_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_25_cast_fp16 = slice_update(begin = concat_195, begin_mask = value_cache_internal_tensor_assign_25_begin_mask_0, end = concat_196, end_mask = value_cache_internal_tensor_assign_25_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_25_squeeze_mask_0, stride = value_cache_internal_tensor_assign_25_stride_0, update = value_states_147_cast_fp16, x = coreml_update_state_103)[name = string("value_cache_internal_tensor_assign_25_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_25_cast_fp16, input = value_cache)[name = string("coreml_update_state_105_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_105 = read_state(input = value_cache)[name = string("coreml_update_state_105")];
            tensor<int32, [4]> var_8792_begin_0 = const()[name = string("op_8792_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_8792_end_0 = const()[name = string("op_8792_end_0"), val = tensor<int32, [4]>([25, 2, 2048, 128])];
            tensor<bool, [4]> var_8792_end_mask_0 = const()[name = string("op_8792_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_8792_cast_fp16 = slice_by_index(begin = var_8792_begin_0, end = var_8792_end_0, end_mask = var_8792_end_mask_0, x = coreml_update_state_104)[name = string("op_8792_cast_fp16")];
            tensor<int32, [2]> tile_48 = const()[name = string("tile_48"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8795_axis_0 = const()[name = string("op_8795_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_8795_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_8795_cast_fp16_1 = split(axis = var_8795_axis_0, split_sizes = tile_48, x = var_8792_cast_fp16)[name = string("op_8795_cast_fp16")];
            tensor<int32, [4]> var_8802_begin_0 = const()[name = string("op_8802_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_8802_end_0 = const()[name = string("op_8802_end_0"), val = tensor<int32, [4]>([25, 2, 2048, 128])];
            tensor<bool, [4]> var_8802_end_mask_0 = const()[name = string("op_8802_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_8802_cast_fp16 = slice_by_index(begin = var_8802_begin_0, end = var_8802_end_0, end_mask = var_8802_end_mask_0, x = coreml_update_state_105)[name = string("op_8802_cast_fp16")];
            tensor<int32, [2]> tile_49 = const()[name = string("tile_49"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8805_axis_0 = const()[name = string("op_8805_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_8805_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_8805_cast_fp16_1 = split(axis = var_8805_axis_0, split_sizes = tile_49, x = var_8802_cast_fp16)[name = string("op_8805_cast_fp16")];
            tensor<int32, [2]> var_8808_split_sizes_0 = const()[name = string("op_8808_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_8808_axis_0 = const()[name = string("op_8808_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_8808_0, tensor<fp16, [1, 8, 128, 8]> var_8808_1 = split(axis = var_8808_axis_0, split_sizes = var_8808_split_sizes_0, x = query_states_147_cast_fp16)[name = string("op_8808")];
            bool attn_weights_385_transpose_x_0 = const()[name = string("attn_weights_385_transpose_x_0"), val = bool(false)];
            bool attn_weights_385_transpose_y_0 = const()[name = string("attn_weights_385_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_385_cast_fp16 = matmul(transpose_x = attn_weights_385_transpose_x_0, transpose_y = attn_weights_385_transpose_y_0, x = var_8795_cast_fp16_0, y = var_8808_0)[name = string("attn_weights_385_cast_fp16")];
            fp16 var_8811_to_fp16 = const()[name = string("op_8811_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_387_cast_fp16 = mul(x = attn_weights_385_cast_fp16, y = var_8811_to_fp16)[name = string("attn_weights_387_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_389_cast_fp16 = add(x = attn_weights_387_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_389_cast_fp16")];
            int32 var_8815 = const()[name = string("op_8815"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_391_cast_fp16 = softmax(axis = var_8815, x = attn_weights_389_cast_fp16)[name = string("attn_weights_391_cast_fp16")];
            bool var_8821_transpose_x_1 = const()[name = string("op_8821_transpose_x_1"), val = bool(true)];
            bool var_8821_transpose_y_1 = const()[name = string("op_8821_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_8821_cast_fp16 = matmul(transpose_x = var_8821_transpose_x_1, transpose_y = var_8821_transpose_y_1, x = attn_weights_391_cast_fp16, y = var_8805_cast_fp16_0)[name = string("op_8821_cast_fp16")];
            bool attn_weights_393_transpose_x_0 = const()[name = string("attn_weights_393_transpose_x_0"), val = bool(false)];
            bool attn_weights_393_transpose_y_0 = const()[name = string("attn_weights_393_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_393_cast_fp16 = matmul(transpose_x = attn_weights_393_transpose_x_0, transpose_y = attn_weights_393_transpose_y_0, x = var_8795_cast_fp16_1, y = var_8808_1)[name = string("attn_weights_393_cast_fp16")];
            fp16 var_8823_to_fp16 = const()[name = string("op_8823_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_395_cast_fp16 = mul(x = attn_weights_393_cast_fp16, y = var_8823_to_fp16)[name = string("attn_weights_395_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_397_cast_fp16 = add(x = attn_weights_395_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_397_cast_fp16")];
            int32 var_8827 = const()[name = string("op_8827"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_399_cast_fp16 = softmax(axis = var_8827, x = attn_weights_397_cast_fp16)[name = string("attn_weights_399_cast_fp16")];
            bool attn_output_193_transpose_x_1 = const()[name = string("attn_output_193_transpose_x_1"), val = bool(true)];
            bool attn_output_193_transpose_y_1 = const()[name = string("attn_output_193_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_193_cast_fp16 = matmul(transpose_x = attn_output_193_transpose_x_1, transpose_y = attn_output_193_transpose_y_1, x = attn_weights_399_cast_fp16, y = var_8805_cast_fp16_1)[name = string("attn_output_193_cast_fp16")];
            int32 var_8835 = const()[name = string("op_8835"), val = int32(1)];
            bool attn_output_195_interleave_0 = const()[name = string("attn_output_195_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_195_cast_fp16 = concat(axis = var_8835, interleave = attn_output_195_interleave_0, values = (var_8821_cast_fp16, attn_output_193_cast_fp16))[name = string("attn_output_195_cast_fp16")];
            tensor<int32, [4]> var_8839_perm_0 = const()[name = string("op_8839_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_8844 = const()[name = string("op_8844"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_8839_cast_fp16 = transpose(perm = var_8839_perm_0, x = attn_output_195_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_199_cast_fp16 = reshape(shape = var_8844, x = var_8839_cast_fp16)[name = string("attn_output_199_cast_fp16")];
            string hidden_states_243_pad_type_0 = const()[name = string("hidden_states_243_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_243_strides_0 = const()[name = string("hidden_states_243_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_243_pad_0 = const()[name = string("hidden_states_243_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_243_dilations_0 = const()[name = string("hidden_states_243_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_243_groups_0 = const()[name = string("hidden_states_243_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_243_cast_fp16 = conv(dilations = hidden_states_243_dilations_0, groups = hidden_states_243_groups_0, pad = hidden_states_243_pad_0, pad_type = hidden_states_243_pad_type_0, strides = hidden_states_243_strides_0, weight = layers_24_self_attn_o_proj_weight_cast_fp16, x = attn_output_199_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_245_cast_fp16 = add(x = hidden_states_239_cast_fp16, y = hidden_states_243_cast_fp16)[name = string("hidden_states_245_cast_fp16")];
            int32 var_8870 = const()[name = string("op_8870"), val = int32(1)];
            fp16 const_301_promoted_to_fp16 = const()[name = string("const_301_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_8872_cast_fp16 = mul(x = hidden_states_245_cast_fp16, y = const_301_promoted_to_fp16)[name = string("op_8872_cast_fp16")];
            bool doubled_197_interleave_0 = const()[name = string("doubled_197_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_197_cast_fp16 = concat(axis = var_8870, interleave = doubled_197_interleave_0, values = (hidden_states_245_cast_fp16, var_8872_cast_fp16))[name = string("doubled_197_cast_fp16")];
            tensor<int32, [1]> out_99_axes_0 = const()[name = string("out_99_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_99_gamma_0_to_fp16 = const()[name = string("out_99_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1532014848)))];
            fp16 var_8882_to_fp16 = const()[name = string("op_8882_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_99_cast_fp16 = layer_norm(axes = out_99_axes_0, epsilon = var_8882_to_fp16, gamma = out_99_gamma_0_to_fp16, x = doubled_197_cast_fp16)[name = string("out_99_cast_fp16")];
            tensor<int32, [2]> var_8893_split_sizes_0 = const()[name = string("op_8893_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_8893_axis_0 = const()[name = string("op_8893_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_8893_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_8893_cast_fp16_1 = split(axis = var_8893_axis_0, split_sizes = var_8893_split_sizes_0, x = out_99_cast_fp16)[name = string("op_8893_cast_fp16")];
            string input_49_pad_type_0 = const()[name = string("input_49_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_49_strides_0 = const()[name = string("input_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_49_pad_0 = const()[name = string("input_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_49_dilations_0 = const()[name = string("input_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_49_groups_0 = const()[name = string("input_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_49_cast_fp16 = conv(dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = layers_24_mlp_gate_proj_weight_cast_fp16, x = var_8893_cast_fp16_0)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_8910_cast_fp16 = silu(x = input_49_cast_fp16)[name = string("op_8910_cast_fp16")];
            string var_8916_pad_type_0 = const()[name = string("op_8916_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_8916_strides_0 = const()[name = string("op_8916_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8916_pad_0 = const()[name = string("op_8916_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8916_dilations_0 = const()[name = string("op_8916_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8916_groups_0 = const()[name = string("op_8916_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_8916_cast_fp16 = conv(dilations = var_8916_dilations_0, groups = var_8916_groups_0, pad = var_8916_pad_0, pad_type = var_8916_pad_type_0, strides = var_8916_strides_0, weight = layers_24_mlp_up_proj_weight_cast_fp16, x = var_8893_cast_fp16_0)[name = string("op_8916_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_249_cast_fp16 = mul(x = var_8910_cast_fp16, y = var_8916_cast_fp16)[name = string("x_249_cast_fp16")];
            string hidden_states_247_pad_type_0 = const()[name = string("hidden_states_247_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_247_strides_0 = const()[name = string("hidden_states_247_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_247_pad_0 = const()[name = string("hidden_states_247_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_247_dilations_0 = const()[name = string("hidden_states_247_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_247_groups_0 = const()[name = string("hidden_states_247_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_247_cast_fp16 = conv(dilations = hidden_states_247_dilations_0, groups = hidden_states_247_groups_0, pad = hidden_states_247_pad_0, pad_type = hidden_states_247_pad_type_0, strides = hidden_states_247_strides_0, weight = layers_24_mlp_down_proj_weight_cast_fp16, x = x_249_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_249_cast_fp16 = add(x = hidden_states_245_cast_fp16, y = hidden_states_247_cast_fp16)[name = string("hidden_states_249_cast_fp16")];
            int32 var_8932 = const()[name = string("op_8932"), val = int32(1)];
            fp16 const_303_promoted_to_fp16 = const()[name = string("const_303_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_8934_cast_fp16 = mul(x = hidden_states_249_cast_fp16, y = const_303_promoted_to_fp16)[name = string("op_8934_cast_fp16")];
            bool doubled_201_interleave_0 = const()[name = string("doubled_201_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_201_cast_fp16 = concat(axis = var_8932, interleave = doubled_201_interleave_0, values = (hidden_states_249_cast_fp16, var_8934_cast_fp16))[name = string("doubled_201_cast_fp16")];
            tensor<int32, [1]> out_101_axes_0 = const()[name = string("out_101_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_101_gamma_0_to_fp16 = const()[name = string("out_101_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1532023104)))];
            fp16 var_8944_to_fp16 = const()[name = string("op_8944_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_101_cast_fp16 = layer_norm(axes = out_101_axes_0, epsilon = var_8944_to_fp16, gamma = out_101_gamma_0_to_fp16, x = doubled_201_cast_fp16)[name = string("out_101_cast_fp16")];
            tensor<int32, [2]> var_8955_split_sizes_0 = const()[name = string("op_8955_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_8955_axis_0 = const()[name = string("op_8955_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_8955_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_8955_cast_fp16_1 = split(axis = var_8955_axis_0, split_sizes = var_8955_split_sizes_0, x = out_101_cast_fp16)[name = string("op_8955_cast_fp16")];
            string query_states_151_pad_type_0 = const()[name = string("query_states_151_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_151_strides_0 = const()[name = string("query_states_151_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_151_pad_0 = const()[name = string("query_states_151_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_151_dilations_0 = const()[name = string("query_states_151_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_151_groups_0 = const()[name = string("query_states_151_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_151_cast_fp16 = conv(dilations = query_states_151_dilations_0, groups = query_states_151_groups_0, pad = query_states_151_pad_0, pad_type = query_states_151_pad_type_0, strides = query_states_151_strides_0, weight = layers_25_self_attn_q_proj_weight_cast_fp16, x = var_8955_cast_fp16_0)[name = string("query_states_151_cast_fp16")];
            string key_states_251_pad_type_0 = const()[name = string("key_states_251_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_251_strides_0 = const()[name = string("key_states_251_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_251_pad_0 = const()[name = string("key_states_251_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_251_dilations_0 = const()[name = string("key_states_251_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_251_groups_0 = const()[name = string("key_states_251_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_251_cast_fp16 = conv(dilations = key_states_251_dilations_0, groups = key_states_251_groups_0, pad = key_states_251_pad_0, pad_type = key_states_251_pad_type_0, strides = key_states_251_strides_0, weight = layers_25_self_attn_k_proj_weight_cast_fp16, x = var_8955_cast_fp16_0)[name = string("key_states_251_cast_fp16")];
            string value_states_151_pad_type_0 = const()[name = string("value_states_151_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_151_strides_0 = const()[name = string("value_states_151_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_151_pad_0 = const()[name = string("value_states_151_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_151_dilations_0 = const()[name = string("value_states_151_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_151_groups_0 = const()[name = string("value_states_151_groups_0"), val = int32(1)];
            tensor<fp16, [256, 2048, 1, 1]> layers_25_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_25_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1532031360)))];
            tensor<fp16, [1, 256, 1, 8]> value_states_151_cast_fp16 = conv(dilations = value_states_151_dilations_0, groups = value_states_151_groups_0, pad = value_states_151_pad_0, pad_type = value_states_151_pad_type_0, strides = value_states_151_strides_0, weight = layers_25_self_attn_v_proj_weight_to_fp16, x = var_8955_cast_fp16_0)[name = string("value_states_151_cast_fp16")];
            tensor<int32, [4]> var_9007 = const()[name = string("op_9007"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_251_cast_fp16 = reshape(shape = var_9007, x = query_states_151_cast_fp16)[name = string("x_251_cast_fp16")];
            tensor<int32, [4]> var_9011 = const()[name = string("op_9011"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_9012_cast_fp16 = reshape(shape = var_9011, x = key_states_251_cast_fp16)[name = string("op_9012_cast_fp16")];
            tensor<int32, [4]> var_9018 = const()[name = string("op_9018"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_9019_cast_fp16 = reshape(shape = var_9018, x = value_states_151_cast_fp16)[name = string("op_9019_cast_fp16")];
            tensor<int32, [4]> value_states_153_perm_0 = const()[name = string("value_states_153_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_9023_cast_fp16 = mul(x = x_251_cast_fp16, y = var_859_cast_fp16)[name = string("op_9023_cast_fp16")];
            tensor<int32, [2]> var_9024_split_sizes_0 = const()[name = string("op_9024_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_9024_axis_0 = const()[name = string("op_9024_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_9024_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_9024_cast_fp16_1 = split(axis = var_9024_axis_0, split_sizes = var_9024_split_sizes_0, x = x_251_cast_fp16)[name = string("op_9024_cast_fp16")];
            fp16 const_307_promoted_to_fp16 = const()[name = string("const_307_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_9026_cast_fp16 = mul(x = var_9024_cast_fp16_1, y = const_307_promoted_to_fp16)[name = string("op_9026_cast_fp16")];
            int32 var_9028 = const()[name = string("op_9028"), val = int32(-2)];
            bool var_9029_interleave_0 = const()[name = string("op_9029_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_9029_cast_fp16 = concat(axis = var_9028, interleave = var_9029_interleave_0, values = (var_9026_cast_fp16, var_9024_cast_fp16_0))[name = string("op_9029_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_9030_cast_fp16 = mul(x = var_9029_cast_fp16, y = var_866_cast_fp16)[name = string("op_9030_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_153_cast_fp16 = add(x = var_9023_cast_fp16, y = var_9030_cast_fp16)[name = string("query_states_153_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_9036_cast_fp16 = mul(x = var_9012_cast_fp16, y = var_859_cast_fp16)[name = string("op_9036_cast_fp16")];
            tensor<int32, [2]> var_9037_split_sizes_0 = const()[name = string("op_9037_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_9037_axis_0 = const()[name = string("op_9037_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_9037_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_9037_cast_fp16_1 = split(axis = var_9037_axis_0, split_sizes = var_9037_split_sizes_0, x = var_9012_cast_fp16)[name = string("op_9037_cast_fp16")];
            fp16 const_308_promoted_to_fp16 = const()[name = string("const_308_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_9039_cast_fp16 = mul(x = var_9037_cast_fp16_1, y = const_308_promoted_to_fp16)[name = string("op_9039_cast_fp16")];
            int32 var_9041 = const()[name = string("op_9041"), val = int32(-2)];
            bool var_9042_interleave_0 = const()[name = string("op_9042_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_9042_cast_fp16 = concat(axis = var_9041, interleave = var_9042_interleave_0, values = (var_9039_cast_fp16, var_9037_cast_fp16_0))[name = string("op_9042_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_9043_cast_fp16 = mul(x = var_9042_cast_fp16, y = var_866_cast_fp16)[name = string("op_9043_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_255_cast_fp16 = add(x = var_9036_cast_fp16, y = var_9043_cast_fp16)[name = string("key_states_255_cast_fp16")];
            tensor<int32, [4]> key_states_257_perm_0 = const()[name = string("key_states_257_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_301 = const()[name = string("expand_dims_301"), val = tensor<int32, [1]>([25])];
            tensor<int32, [1]> expand_dims_302 = const()[name = string("expand_dims_302"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_304 = const()[name = string("expand_dims_304"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_305 = const()[name = string("expand_dims_305"), val = tensor<int32, [1]>([26])];
            int32 concat_203_axis_0 = const()[name = string("concat_203_axis_0"), val = int32(0)];
            bool concat_203_interleave_0 = const()[name = string("concat_203_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_203 = concat(axis = concat_203_axis_0, interleave = concat_203_interleave_0, values = (expand_dims_301, expand_dims_302, position_id, expand_dims_304))[name = string("concat_203")];
            tensor<int32, [1]> concat_204_values1_0 = const()[name = string("concat_204_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_204_values3_0 = const()[name = string("concat_204_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_204_axis_0 = const()[name = string("concat_204_axis_0"), val = int32(0)];
            bool concat_204_interleave_0 = const()[name = string("concat_204_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_204 = concat(axis = concat_204_axis_0, interleave = concat_204_interleave_0, values = (expand_dims_305, concat_204_values1_0, cache_position_end, concat_204_values3_0))[name = string("concat_204")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_26_stride_0 = const()[name = string("key_cache_internal_tensor_assign_26_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_26_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_26_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_26_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_26_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_26_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_26_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_257_cast_fp16 = transpose(perm = key_states_257_perm_0, x = key_states_255_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_26_cast_fp16 = slice_update(begin = concat_203, begin_mask = key_cache_internal_tensor_assign_26_begin_mask_0, end = concat_204, end_mask = key_cache_internal_tensor_assign_26_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_26_squeeze_mask_0, stride = key_cache_internal_tensor_assign_26_stride_0, update = key_states_257_cast_fp16, x = coreml_update_state_104)[name = string("key_cache_internal_tensor_assign_26_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_26_cast_fp16, input = key_cache)[name = string("coreml_update_state_106_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_106 = read_state(input = key_cache)[name = string("coreml_update_state_106")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_26_stride_0 = const()[name = string("value_cache_internal_tensor_assign_26_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_26_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_26_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_26_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_26_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_26_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_26_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_153_cast_fp16 = transpose(perm = value_states_153_perm_0, x = var_9019_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_26_cast_fp16 = slice_update(begin = concat_203, begin_mask = value_cache_internal_tensor_assign_26_begin_mask_0, end = concat_204, end_mask = value_cache_internal_tensor_assign_26_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_26_squeeze_mask_0, stride = value_cache_internal_tensor_assign_26_stride_0, update = value_states_153_cast_fp16, x = coreml_update_state_105)[name = string("value_cache_internal_tensor_assign_26_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_26_cast_fp16, input = value_cache)[name = string("coreml_update_state_107_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_107 = read_state(input = value_cache)[name = string("coreml_update_state_107")];
            tensor<int32, [4]> var_9113_begin_0 = const()[name = string("op_9113_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9113_end_0 = const()[name = string("op_9113_end_0"), val = tensor<int32, [4]>([26, 2, 2048, 128])];
            tensor<bool, [4]> var_9113_end_mask_0 = const()[name = string("op_9113_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_9113_cast_fp16 = slice_by_index(begin = var_9113_begin_0, end = var_9113_end_0, end_mask = var_9113_end_mask_0, x = coreml_update_state_106)[name = string("op_9113_cast_fp16")];
            tensor<int32, [2]> tile_50 = const()[name = string("tile_50"), val = tensor<int32, [2]>([1, 1])];
            int32 var_9116_axis_0 = const()[name = string("op_9116_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_9116_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_9116_cast_fp16_1 = split(axis = var_9116_axis_0, split_sizes = tile_50, x = var_9113_cast_fp16)[name = string("op_9116_cast_fp16")];
            tensor<int32, [4]> var_9123_begin_0 = const()[name = string("op_9123_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9123_end_0 = const()[name = string("op_9123_end_0"), val = tensor<int32, [4]>([26, 2, 2048, 128])];
            tensor<bool, [4]> var_9123_end_mask_0 = const()[name = string("op_9123_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_9123_cast_fp16 = slice_by_index(begin = var_9123_begin_0, end = var_9123_end_0, end_mask = var_9123_end_mask_0, x = coreml_update_state_107)[name = string("op_9123_cast_fp16")];
            tensor<int32, [2]> tile_51 = const()[name = string("tile_51"), val = tensor<int32, [2]>([1, 1])];
            int32 var_9126_axis_0 = const()[name = string("op_9126_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_9126_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_9126_cast_fp16_1 = split(axis = var_9126_axis_0, split_sizes = tile_51, x = var_9123_cast_fp16)[name = string("op_9126_cast_fp16")];
            tensor<int32, [2]> var_9129_split_sizes_0 = const()[name = string("op_9129_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_9129_axis_0 = const()[name = string("op_9129_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_9129_0, tensor<fp16, [1, 8, 128, 8]> var_9129_1 = split(axis = var_9129_axis_0, split_sizes = var_9129_split_sizes_0, x = query_states_153_cast_fp16)[name = string("op_9129")];
            bool attn_weights_401_transpose_x_0 = const()[name = string("attn_weights_401_transpose_x_0"), val = bool(false)];
            bool attn_weights_401_transpose_y_0 = const()[name = string("attn_weights_401_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_401_cast_fp16 = matmul(transpose_x = attn_weights_401_transpose_x_0, transpose_y = attn_weights_401_transpose_y_0, x = var_9116_cast_fp16_0, y = var_9129_0)[name = string("attn_weights_401_cast_fp16")];
            fp16 var_9132_to_fp16 = const()[name = string("op_9132_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_403_cast_fp16 = mul(x = attn_weights_401_cast_fp16, y = var_9132_to_fp16)[name = string("attn_weights_403_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_405_cast_fp16 = add(x = attn_weights_403_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_405_cast_fp16")];
            int32 var_9136 = const()[name = string("op_9136"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_407_cast_fp16 = softmax(axis = var_9136, x = attn_weights_405_cast_fp16)[name = string("attn_weights_407_cast_fp16")];
            bool var_9142_transpose_x_1 = const()[name = string("op_9142_transpose_x_1"), val = bool(true)];
            bool var_9142_transpose_y_1 = const()[name = string("op_9142_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_9142_cast_fp16 = matmul(transpose_x = var_9142_transpose_x_1, transpose_y = var_9142_transpose_y_1, x = attn_weights_407_cast_fp16, y = var_9126_cast_fp16_0)[name = string("op_9142_cast_fp16")];
            bool attn_weights_409_transpose_x_0 = const()[name = string("attn_weights_409_transpose_x_0"), val = bool(false)];
            bool attn_weights_409_transpose_y_0 = const()[name = string("attn_weights_409_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_409_cast_fp16 = matmul(transpose_x = attn_weights_409_transpose_x_0, transpose_y = attn_weights_409_transpose_y_0, x = var_9116_cast_fp16_1, y = var_9129_1)[name = string("attn_weights_409_cast_fp16")];
            fp16 var_9144_to_fp16 = const()[name = string("op_9144_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_411_cast_fp16 = mul(x = attn_weights_409_cast_fp16, y = var_9144_to_fp16)[name = string("attn_weights_411_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_413_cast_fp16 = add(x = attn_weights_411_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_413_cast_fp16")];
            int32 var_9148 = const()[name = string("op_9148"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_415_cast_fp16 = softmax(axis = var_9148, x = attn_weights_413_cast_fp16)[name = string("attn_weights_415_cast_fp16")];
            bool attn_output_201_transpose_x_1 = const()[name = string("attn_output_201_transpose_x_1"), val = bool(true)];
            bool attn_output_201_transpose_y_1 = const()[name = string("attn_output_201_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_201_cast_fp16 = matmul(transpose_x = attn_output_201_transpose_x_1, transpose_y = attn_output_201_transpose_y_1, x = attn_weights_415_cast_fp16, y = var_9126_cast_fp16_1)[name = string("attn_output_201_cast_fp16")];
            int32 var_9156 = const()[name = string("op_9156"), val = int32(1)];
            bool attn_output_203_interleave_0 = const()[name = string("attn_output_203_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_203_cast_fp16 = concat(axis = var_9156, interleave = attn_output_203_interleave_0, values = (var_9142_cast_fp16, attn_output_201_cast_fp16))[name = string("attn_output_203_cast_fp16")];
            tensor<int32, [4]> var_9160_perm_0 = const()[name = string("op_9160_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_9165 = const()[name = string("op_9165"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_9160_cast_fp16 = transpose(perm = var_9160_perm_0, x = attn_output_203_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_207_cast_fp16 = reshape(shape = var_9165, x = var_9160_cast_fp16)[name = string("attn_output_207_cast_fp16")];
            string hidden_states_253_pad_type_0 = const()[name = string("hidden_states_253_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_253_strides_0 = const()[name = string("hidden_states_253_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_253_pad_0 = const()[name = string("hidden_states_253_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_253_dilations_0 = const()[name = string("hidden_states_253_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_253_groups_0 = const()[name = string("hidden_states_253_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_253_cast_fp16 = conv(dilations = hidden_states_253_dilations_0, groups = hidden_states_253_groups_0, pad = hidden_states_253_pad_0, pad_type = hidden_states_253_pad_type_0, strides = hidden_states_253_strides_0, weight = layers_25_self_attn_o_proj_weight_cast_fp16, x = attn_output_207_cast_fp16)[name = string("hidden_states_253_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_255_cast_fp16 = add(x = hidden_states_249_cast_fp16, y = hidden_states_253_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            int32 var_9191 = const()[name = string("op_9191"), val = int32(1)];
            fp16 const_313_promoted_to_fp16 = const()[name = string("const_313_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_9193_cast_fp16 = mul(x = hidden_states_255_cast_fp16, y = const_313_promoted_to_fp16)[name = string("op_9193_cast_fp16")];
            bool doubled_205_interleave_0 = const()[name = string("doubled_205_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_205_cast_fp16 = concat(axis = var_9191, interleave = doubled_205_interleave_0, values = (hidden_states_255_cast_fp16, var_9193_cast_fp16))[name = string("doubled_205_cast_fp16")];
            tensor<int32, [1]> out_103_axes_0 = const()[name = string("out_103_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_103_gamma_0_to_fp16 = const()[name = string("out_103_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1533080000)))];
            fp16 var_9203_to_fp16 = const()[name = string("op_9203_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_103_cast_fp16 = layer_norm(axes = out_103_axes_0, epsilon = var_9203_to_fp16, gamma = out_103_gamma_0_to_fp16, x = doubled_205_cast_fp16)[name = string("out_103_cast_fp16")];
            tensor<int32, [2]> var_9214_split_sizes_0 = const()[name = string("op_9214_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_9214_axis_0 = const()[name = string("op_9214_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_9214_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_9214_cast_fp16_1 = split(axis = var_9214_axis_0, split_sizes = var_9214_split_sizes_0, x = out_103_cast_fp16)[name = string("op_9214_cast_fp16")];
            string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = layers_25_mlp_gate_proj_weight_cast_fp16, x = var_9214_cast_fp16_0)[name = string("input_51_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_9231_cast_fp16 = silu(x = input_51_cast_fp16)[name = string("op_9231_cast_fp16")];
            string var_9237_pad_type_0 = const()[name = string("op_9237_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_9237_strides_0 = const()[name = string("op_9237_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9237_pad_0 = const()[name = string("op_9237_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9237_dilations_0 = const()[name = string("op_9237_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_9237_groups_0 = const()[name = string("op_9237_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> var_9237_cast_fp16 = conv(dilations = var_9237_dilations_0, groups = var_9237_groups_0, pad = var_9237_pad_0, pad_type = var_9237_pad_type_0, strides = var_9237_strides_0, weight = layers_25_mlp_up_proj_weight_cast_fp16, x = var_9214_cast_fp16_0)[name = string("op_9237_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_259_cast_fp16 = mul(x = var_9231_cast_fp16, y = var_9237_cast_fp16)[name = string("x_259_cast_fp16")];
            string hidden_states_257_pad_type_0 = const()[name = string("hidden_states_257_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_257_strides_0 = const()[name = string("hidden_states_257_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_257_pad_0 = const()[name = string("hidden_states_257_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_257_dilations_0 = const()[name = string("hidden_states_257_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_257_groups_0 = const()[name = string("hidden_states_257_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 6144, 1, 1]> layers_25_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_25_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1533088256)))];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_257_cast_fp16 = conv(dilations = hidden_states_257_dilations_0, groups = hidden_states_257_groups_0, pad = hidden_states_257_pad_0, pad_type = hidden_states_257_pad_type_0, strides = hidden_states_257_strides_0, weight = layers_25_mlp_down_proj_weight_to_fp16, x = x_259_cast_fp16)[name = string("hidden_states_257_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_259_cast_fp16 = add(x = hidden_states_255_cast_fp16, y = hidden_states_257_cast_fp16)[name = string("hidden_states_259_cast_fp16")];
            int32 var_9253 = const()[name = string("op_9253"), val = int32(1)];
            fp16 const_315_promoted_to_fp16 = const()[name = string("const_315_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_9255_cast_fp16 = mul(x = hidden_states_259_cast_fp16, y = const_315_promoted_to_fp16)[name = string("op_9255_cast_fp16")];
            bool doubled_209_interleave_0 = const()[name = string("doubled_209_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_209_cast_fp16 = concat(axis = var_9253, interleave = doubled_209_interleave_0, values = (hidden_states_259_cast_fp16, var_9255_cast_fp16))[name = string("doubled_209_cast_fp16")];
            tensor<int32, [1]> out_105_axes_0 = const()[name = string("out_105_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_105_gamma_0_to_fp16 = const()[name = string("out_105_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1558254144)))];
            fp16 var_9265_to_fp16 = const()[name = string("op_9265_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_105_cast_fp16 = layer_norm(axes = out_105_axes_0, epsilon = var_9265_to_fp16, gamma = out_105_gamma_0_to_fp16, x = doubled_209_cast_fp16)[name = string("out_105_cast_fp16")];
            tensor<int32, [2]> var_9276_split_sizes_0 = const()[name = string("op_9276_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_9276_axis_0 = const()[name = string("op_9276_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_9276_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_9276_cast_fp16_1 = split(axis = var_9276_axis_0, split_sizes = var_9276_split_sizes_0, x = out_105_cast_fp16)[name = string("op_9276_cast_fp16")];
            string query_states_157_pad_type_0 = const()[name = string("query_states_157_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_157_strides_0 = const()[name = string("query_states_157_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_157_pad_0 = const()[name = string("query_states_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_157_dilations_0 = const()[name = string("query_states_157_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_157_groups_0 = const()[name = string("query_states_157_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> query_states_157_cast_fp16 = conv(dilations = query_states_157_dilations_0, groups = query_states_157_groups_0, pad = query_states_157_pad_0, pad_type = query_states_157_pad_type_0, strides = query_states_157_strides_0, weight = layers_26_self_attn_q_proj_weight_cast_fp16, x = var_9276_cast_fp16_0)[name = string("query_states_157_cast_fp16")];
            string key_states_261_pad_type_0 = const()[name = string("key_states_261_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_261_strides_0 = const()[name = string("key_states_261_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_261_pad_0 = const()[name = string("key_states_261_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_261_dilations_0 = const()[name = string("key_states_261_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_261_groups_0 = const()[name = string("key_states_261_groups_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, 8]> key_states_261_cast_fp16 = conv(dilations = key_states_261_dilations_0, groups = key_states_261_groups_0, pad = key_states_261_pad_0, pad_type = key_states_261_pad_type_0, strides = key_states_261_strides_0, weight = layers_26_self_attn_k_proj_weight_cast_fp16, x = var_9276_cast_fp16_0)[name = string("key_states_261_cast_fp16")];
            string value_states_157_pad_type_0 = const()[name = string("value_states_157_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_157_strides_0 = const()[name = string("value_states_157_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_157_pad_0 = const()[name = string("value_states_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_157_dilations_0 = const()[name = string("value_states_157_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_157_groups_0 = const()[name = string("value_states_157_groups_0"), val = int32(1)];
            tensor<fp16, [256, 2048, 1, 1]> layers_26_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_26_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1558262400)))];
            tensor<fp16, [1, 256, 1, 8]> value_states_157_cast_fp16 = conv(dilations = value_states_157_dilations_0, groups = value_states_157_groups_0, pad = value_states_157_pad_0, pad_type = value_states_157_pad_type_0, strides = value_states_157_strides_0, weight = layers_26_self_attn_v_proj_weight_to_fp16, x = var_9276_cast_fp16_0)[name = string("value_states_157_cast_fp16")];
            tensor<int32, [4]> var_9328 = const()[name = string("op_9328"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_261_cast_fp16 = reshape(shape = var_9328, x = query_states_157_cast_fp16)[name = string("x_261_cast_fp16")];
            tensor<int32, [4]> var_9332 = const()[name = string("op_9332"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_9333_cast_fp16 = reshape(shape = var_9332, x = key_states_261_cast_fp16)[name = string("op_9333_cast_fp16")];
            tensor<int32, [4]> var_9339 = const()[name = string("op_9339"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_9340_cast_fp16 = reshape(shape = var_9339, x = value_states_157_cast_fp16)[name = string("op_9340_cast_fp16")];
            tensor<int32, [4]> value_states_159_perm_0 = const()[name = string("value_states_159_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_9344_cast_fp16 = mul(x = x_261_cast_fp16, y = var_859_cast_fp16)[name = string("op_9344_cast_fp16")];
            tensor<int32, [2]> var_9345_split_sizes_0 = const()[name = string("op_9345_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_9345_axis_0 = const()[name = string("op_9345_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_9345_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_9345_cast_fp16_1 = split(axis = var_9345_axis_0, split_sizes = var_9345_split_sizes_0, x = x_261_cast_fp16)[name = string("op_9345_cast_fp16")];
            fp16 const_319_promoted_to_fp16 = const()[name = string("const_319_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_9347_cast_fp16 = mul(x = var_9345_cast_fp16_1, y = const_319_promoted_to_fp16)[name = string("op_9347_cast_fp16")];
            int32 var_9349 = const()[name = string("op_9349"), val = int32(-2)];
            bool var_9350_interleave_0 = const()[name = string("op_9350_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_9350_cast_fp16 = concat(axis = var_9349, interleave = var_9350_interleave_0, values = (var_9347_cast_fp16, var_9345_cast_fp16_0))[name = string("op_9350_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_9351_cast_fp16 = mul(x = var_9350_cast_fp16, y = var_866_cast_fp16)[name = string("op_9351_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_159_cast_fp16 = add(x = var_9344_cast_fp16, y = var_9351_cast_fp16)[name = string("query_states_159_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_9357_cast_fp16 = mul(x = var_9333_cast_fp16, y = var_859_cast_fp16)[name = string("op_9357_cast_fp16")];
            tensor<int32, [2]> var_9358_split_sizes_0 = const()[name = string("op_9358_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_9358_axis_0 = const()[name = string("op_9358_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_9358_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_9358_cast_fp16_1 = split(axis = var_9358_axis_0, split_sizes = var_9358_split_sizes_0, x = var_9333_cast_fp16)[name = string("op_9358_cast_fp16")];
            fp16 const_320_promoted_to_fp16 = const()[name = string("const_320_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_9360_cast_fp16 = mul(x = var_9358_cast_fp16_1, y = const_320_promoted_to_fp16)[name = string("op_9360_cast_fp16")];
            int32 var_9362 = const()[name = string("op_9362"), val = int32(-2)];
            bool var_9363_interleave_0 = const()[name = string("op_9363_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_9363_cast_fp16 = concat(axis = var_9362, interleave = var_9363_interleave_0, values = (var_9360_cast_fp16, var_9358_cast_fp16_0))[name = string("op_9363_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_9364_cast_fp16 = mul(x = var_9363_cast_fp16, y = var_866_cast_fp16)[name = string("op_9364_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_265_cast_fp16 = add(x = var_9357_cast_fp16, y = var_9364_cast_fp16)[name = string("key_states_265_cast_fp16")];
            tensor<int32, [4]> key_states_267_perm_0 = const()[name = string("key_states_267_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_313 = const()[name = string("expand_dims_313"), val = tensor<int32, [1]>([26])];
            tensor<int32, [1]> expand_dims_314 = const()[name = string("expand_dims_314"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_316 = const()[name = string("expand_dims_316"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_317 = const()[name = string("expand_dims_317"), val = tensor<int32, [1]>([27])];
            int32 concat_211_axis_0 = const()[name = string("concat_211_axis_0"), val = int32(0)];
            bool concat_211_interleave_0 = const()[name = string("concat_211_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_211 = concat(axis = concat_211_axis_0, interleave = concat_211_interleave_0, values = (expand_dims_313, expand_dims_314, position_id, expand_dims_316))[name = string("concat_211")];
            tensor<int32, [1]> concat_212_values1_0 = const()[name = string("concat_212_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_212_values3_0 = const()[name = string("concat_212_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_212_axis_0 = const()[name = string("concat_212_axis_0"), val = int32(0)];
            bool concat_212_interleave_0 = const()[name = string("concat_212_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_212 = concat(axis = concat_212_axis_0, interleave = concat_212_interleave_0, values = (expand_dims_317, concat_212_values1_0, cache_position_end, concat_212_values3_0))[name = string("concat_212")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_27_stride_0 = const()[name = string("key_cache_internal_tensor_assign_27_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_27_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_27_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_27_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_27_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_27_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_27_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_267_cast_fp16 = transpose(perm = key_states_267_perm_0, x = key_states_265_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_27_cast_fp16 = slice_update(begin = concat_211, begin_mask = key_cache_internal_tensor_assign_27_begin_mask_0, end = concat_212, end_mask = key_cache_internal_tensor_assign_27_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_27_squeeze_mask_0, stride = key_cache_internal_tensor_assign_27_stride_0, update = key_states_267_cast_fp16, x = coreml_update_state_106)[name = string("key_cache_internal_tensor_assign_27_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_27_cast_fp16, input = key_cache)[name = string("coreml_update_state_108_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_108 = read_state(input = key_cache)[name = string("coreml_update_state_108")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_27_stride_0 = const()[name = string("value_cache_internal_tensor_assign_27_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_27_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_27_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_27_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_27_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_27_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_27_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_159_cast_fp16 = transpose(perm = value_states_159_perm_0, x = var_9340_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_27_cast_fp16 = slice_update(begin = concat_211, begin_mask = value_cache_internal_tensor_assign_27_begin_mask_0, end = concat_212, end_mask = value_cache_internal_tensor_assign_27_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_27_squeeze_mask_0, stride = value_cache_internal_tensor_assign_27_stride_0, update = value_states_159_cast_fp16, x = coreml_update_state_107)[name = string("value_cache_internal_tensor_assign_27_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_27_cast_fp16, input = value_cache)[name = string("coreml_update_state_109_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_109 = read_state(input = value_cache)[name = string("coreml_update_state_109")];
            tensor<int32, [4]> var_9434_begin_0 = const()[name = string("op_9434_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_9434_end_0 = const()[name = string("op_9434_end_0"), val = tensor<int32, [4]>([27, 2, 2048, 128])];
            tensor<bool, [4]> var_9434_end_mask_0 = const()[name = string("op_9434_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_9434_cast_fp16 = slice_by_index(begin = var_9434_begin_0, end = var_9434_end_0, end_mask = var_9434_end_mask_0, x = coreml_update_state_108)[name = string("op_9434_cast_fp16")];
            tensor<int32, [2]> tile_52 = const()[name = string("tile_52"), val = tensor<int32, [2]>([1, 1])];
            int32 var_9437_axis_0 = const()[name = string("op_9437_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_9437_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_9437_cast_fp16_1 = split(axis = var_9437_axis_0, split_sizes = tile_52, x = var_9434_cast_fp16)[name = string("op_9437_cast_fp16")];
            tensor<int32, [4]> var_9444_begin_0 = const()[name = string("op_9444_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_9444_end_0 = const()[name = string("op_9444_end_0"), val = tensor<int32, [4]>([27, 2, 2048, 128])];
            tensor<bool, [4]> var_9444_end_mask_0 = const()[name = string("op_9444_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_9444_cast_fp16 = slice_by_index(begin = var_9444_begin_0, end = var_9444_end_0, end_mask = var_9444_end_mask_0, x = coreml_update_state_109)[name = string("op_9444_cast_fp16")];
            tensor<int32, [2]> tile_53 = const()[name = string("tile_53"), val = tensor<int32, [2]>([1, 1])];
            int32 var_9447_axis_0 = const()[name = string("op_9447_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_9447_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_9447_cast_fp16_1 = split(axis = var_9447_axis_0, split_sizes = tile_53, x = var_9444_cast_fp16)[name = string("op_9447_cast_fp16")];
            tensor<int32, [2]> var_9450_split_sizes_0 = const()[name = string("op_9450_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_9450_axis_0 = const()[name = string("op_9450_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_9450_0, tensor<fp16, [1, 8, 128, 8]> var_9450_1 = split(axis = var_9450_axis_0, split_sizes = var_9450_split_sizes_0, x = query_states_159_cast_fp16)[name = string("op_9450")];
            bool attn_weights_417_transpose_x_0 = const()[name = string("attn_weights_417_transpose_x_0"), val = bool(false)];
            bool attn_weights_417_transpose_y_0 = const()[name = string("attn_weights_417_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_417_cast_fp16 = matmul(transpose_x = attn_weights_417_transpose_x_0, transpose_y = attn_weights_417_transpose_y_0, x = var_9437_cast_fp16_0, y = var_9450_0)[name = string("attn_weights_417_cast_fp16")];
            fp16 var_9453_to_fp16 = const()[name = string("op_9453_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_419_cast_fp16 = mul(x = attn_weights_417_cast_fp16, y = var_9453_to_fp16)[name = string("attn_weights_419_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_421_cast_fp16 = add(x = attn_weights_419_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_421_cast_fp16")];
            int32 var_9457 = const()[name = string("op_9457"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_423_cast_fp16 = softmax(axis = var_9457, x = attn_weights_421_cast_fp16)[name = string("attn_weights_423_cast_fp16")];
            bool var_9463_transpose_x_1 = const()[name = string("op_9463_transpose_x_1"), val = bool(true)];
            bool var_9463_transpose_y_1 = const()[name = string("op_9463_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_9463_cast_fp16 = matmul(transpose_x = var_9463_transpose_x_1, transpose_y = var_9463_transpose_y_1, x = attn_weights_423_cast_fp16, y = var_9447_cast_fp16_0)[name = string("op_9463_cast_fp16")];
            bool attn_weights_425_transpose_x_0 = const()[name = string("attn_weights_425_transpose_x_0"), val = bool(false)];
            bool attn_weights_425_transpose_y_0 = const()[name = string("attn_weights_425_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_425_cast_fp16 = matmul(transpose_x = attn_weights_425_transpose_x_0, transpose_y = attn_weights_425_transpose_y_0, x = var_9437_cast_fp16_1, y = var_9450_1)[name = string("attn_weights_425_cast_fp16")];
            fp16 var_9465_to_fp16 = const()[name = string("op_9465_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_427_cast_fp16 = mul(x = attn_weights_425_cast_fp16, y = var_9465_to_fp16)[name = string("attn_weights_427_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_429_cast_fp16 = add(x = attn_weights_427_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_429_cast_fp16")];
            int32 var_9469 = const()[name = string("op_9469"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_431_cast_fp16 = softmax(axis = var_9469, x = attn_weights_429_cast_fp16)[name = string("attn_weights_431_cast_fp16")];
            bool attn_output_209_transpose_x_1 = const()[name = string("attn_output_209_transpose_x_1"), val = bool(true)];
            bool attn_output_209_transpose_y_1 = const()[name = string("attn_output_209_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_209_cast_fp16 = matmul(transpose_x = attn_output_209_transpose_x_1, transpose_y = attn_output_209_transpose_y_1, x = attn_weights_431_cast_fp16, y = var_9447_cast_fp16_1)[name = string("attn_output_209_cast_fp16")];
            int32 var_9477 = const()[name = string("op_9477"), val = int32(1)];
            bool attn_output_211_interleave_0 = const()[name = string("attn_output_211_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_211_cast_fp16 = concat(axis = var_9477, interleave = attn_output_211_interleave_0, values = (var_9463_cast_fp16, attn_output_209_cast_fp16))[name = string("attn_output_211_cast_fp16")];
            tensor<int32, [4]> var_9481_perm_0 = const()[name = string("op_9481_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_9486 = const()[name = string("op_9486"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_9481_cast_fp16 = transpose(perm = var_9481_perm_0, x = attn_output_211_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_215_cast_fp16 = reshape(shape = var_9486, x = var_9481_cast_fp16)[name = string("attn_output_215_cast_fp16")];
            string hidden_states_263_pad_type_0 = const()[name = string("hidden_states_263_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_263_strides_0 = const()[name = string("hidden_states_263_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_263_pad_0 = const()[name = string("hidden_states_263_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_263_dilations_0 = const()[name = string("hidden_states_263_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_263_groups_0 = const()[name = string("hidden_states_263_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_263_cast_fp16 = conv(dilations = hidden_states_263_dilations_0, groups = hidden_states_263_groups_0, pad = hidden_states_263_pad_0, pad_type = hidden_states_263_pad_type_0, strides = hidden_states_263_strides_0, weight = layers_26_self_attn_o_proj_weight_cast_fp16, x = attn_output_215_cast_fp16)[name = string("hidden_states_263_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_265_cast_fp16 = add(x = hidden_states_259_cast_fp16, y = hidden_states_263_cast_fp16)[name = string("hidden_states_265_cast_fp16")];
            int32 var_9512 = const()[name = string("op_9512"), val = int32(1)];
            fp16 const_325_promoted_to_fp16 = const()[name = string("const_325_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_9514_cast_fp16 = mul(x = hidden_states_265_cast_fp16, y = const_325_promoted_to_fp16)[name = string("op_9514_cast_fp16")];
            bool doubled_213_interleave_0 = const()[name = string("doubled_213_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_213_cast_fp16 = concat(axis = var_9512, interleave = doubled_213_interleave_0, values = (hidden_states_265_cast_fp16, var_9514_cast_fp16))[name = string("doubled_213_cast_fp16")];
            tensor<int32, [1]> out_107_axes_0 = const()[name = string("out_107_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_107_gamma_0_to_fp16 = const()[name = string("out_107_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1559311040)))];
            fp16 var_9524_to_fp16 = const()[name = string("op_9524_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_107_cast_fp16 = layer_norm(axes = out_107_axes_0, epsilon = var_9524_to_fp16, gamma = out_107_gamma_0_to_fp16, x = doubled_213_cast_fp16)[name = string("out_107_cast_fp16")];
            tensor<int32, [2]> var_9535_split_sizes_0 = const()[name = string("op_9535_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_9535_axis_0 = const()[name = string("op_9535_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_9535_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_9535_cast_fp16_1 = split(axis = var_9535_axis_0, split_sizes = var_9535_split_sizes_0, x = out_107_cast_fp16)[name = string("op_9535_cast_fp16")];
            string input_53_pad_type_0 = const()[name = string("input_53_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_53_strides_0 = const()[name = string("input_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_53_pad_0 = const()[name = string("input_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_53_dilations_0 = const()[name = string("input_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_53_groups_0 = const()[name = string("input_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 6144, 1, 8]> input_53_cast_fp16 = conv(dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = layers_26_mlp_gate_proj_weight_cast_fp16, x = var_9535_cast_fp16_0)[name = string("input_53_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_9552_cast_fp16 = silu(x = input_53_cast_fp16)[name = string("op_9552_cast_fp16")];
            string var_9558_pad_type_0 = const()[name = string("op_9558_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_9558_strides_0 = const()[name = string("op_9558_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9558_pad_0 = const()[name = string("op_9558_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9558_dilations_0 = const()[name = string("op_9558_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_9558_groups_0 = const()[name = string("op_9558_groups_0"), val = int32(1)];
            tensor<fp16, [6144, 2048, 1, 1]> layers_26_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_26_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1559319296)))];
            tensor<fp16, [1, 6144, 1, 8]> var_9558_cast_fp16 = conv(dilations = var_9558_dilations_0, groups = var_9558_groups_0, pad = var_9558_pad_0, pad_type = var_9558_pad_type_0, strides = var_9558_strides_0, weight = layers_26_mlp_up_proj_weight_to_fp16, x = var_9535_cast_fp16_0)[name = string("op_9558_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_269_cast_fp16 = mul(x = var_9552_cast_fp16, y = var_9558_cast_fp16)[name = string("x_269_cast_fp16")];
            string hidden_states_267_pad_type_0 = const()[name = string("hidden_states_267_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_267_strides_0 = const()[name = string("hidden_states_267_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_267_pad_0 = const()[name = string("hidden_states_267_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_267_dilations_0 = const()[name = string("hidden_states_267_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_267_groups_0 = const()[name = string("hidden_states_267_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 6144, 1, 1]> layers_26_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_26_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1584485184)))];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_267_cast_fp16 = conv(dilations = hidden_states_267_dilations_0, groups = hidden_states_267_groups_0, pad = hidden_states_267_pad_0, pad_type = hidden_states_267_pad_type_0, strides = hidden_states_267_strides_0, weight = layers_26_mlp_down_proj_weight_to_fp16, x = x_269_cast_fp16)[name = string("hidden_states_267_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_269_cast_fp16 = add(x = hidden_states_265_cast_fp16, y = hidden_states_267_cast_fp16)[name = string("hidden_states_269_cast_fp16")];
            int32 var_9574 = const()[name = string("op_9574"), val = int32(1)];
            fp16 const_327_promoted_to_fp16 = const()[name = string("const_327_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_9576_cast_fp16 = mul(x = hidden_states_269_cast_fp16, y = const_327_promoted_to_fp16)[name = string("op_9576_cast_fp16")];
            bool doubled_217_interleave_0 = const()[name = string("doubled_217_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_217_cast_fp16 = concat(axis = var_9574, interleave = doubled_217_interleave_0, values = (hidden_states_269_cast_fp16, var_9576_cast_fp16))[name = string("doubled_217_cast_fp16")];
            tensor<int32, [1]> out_109_axes_0 = const()[name = string("out_109_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_109_gamma_0_to_fp16 = const()[name = string("out_109_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1609651072)))];
            fp16 var_9586_to_fp16 = const()[name = string("op_9586_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_109_cast_fp16 = layer_norm(axes = out_109_axes_0, epsilon = var_9586_to_fp16, gamma = out_109_gamma_0_to_fp16, x = doubled_217_cast_fp16)[name = string("out_109_cast_fp16")];
            tensor<int32, [2]> var_9597_split_sizes_0 = const()[name = string("op_9597_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_9597_axis_0 = const()[name = string("op_9597_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_9597_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_9597_cast_fp16_1 = split(axis = var_9597_axis_0, split_sizes = var_9597_split_sizes_0, x = out_109_cast_fp16)[name = string("op_9597_cast_fp16")];
            string query_states_163_pad_type_0 = const()[name = string("query_states_163_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> query_states_163_strides_0 = const()[name = string("query_states_163_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> query_states_163_pad_0 = const()[name = string("query_states_163_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> query_states_163_dilations_0 = const()[name = string("query_states_163_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 query_states_163_groups_0 = const()[name = string("query_states_163_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 2048, 1, 1]> layers_27_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_27_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1609659328)))];
            tensor<fp16, [1, 2048, 1, 8]> query_states_163_cast_fp16 = conv(dilations = query_states_163_dilations_0, groups = query_states_163_groups_0, pad = query_states_163_pad_0, pad_type = query_states_163_pad_type_0, strides = query_states_163_strides_0, weight = layers_27_self_attn_q_proj_weight_to_fp16, x = var_9597_cast_fp16_0)[name = string("query_states_163_cast_fp16")];
            string key_states_271_pad_type_0 = const()[name = string("key_states_271_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> key_states_271_strides_0 = const()[name = string("key_states_271_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> key_states_271_pad_0 = const()[name = string("key_states_271_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> key_states_271_dilations_0 = const()[name = string("key_states_271_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 key_states_271_groups_0 = const()[name = string("key_states_271_groups_0"), val = int32(1)];
            tensor<fp16, [256, 2048, 1, 1]> layers_27_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_27_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1618048000)))];
            tensor<fp16, [1, 256, 1, 8]> key_states_271_cast_fp16 = conv(dilations = key_states_271_dilations_0, groups = key_states_271_groups_0, pad = key_states_271_pad_0, pad_type = key_states_271_pad_type_0, strides = key_states_271_strides_0, weight = layers_27_self_attn_k_proj_weight_to_fp16, x = var_9597_cast_fp16_0)[name = string("key_states_271_cast_fp16")];
            string value_states_163_pad_type_0 = const()[name = string("value_states_163_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> value_states_163_strides_0 = const()[name = string("value_states_163_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> value_states_163_pad_0 = const()[name = string("value_states_163_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> value_states_163_dilations_0 = const()[name = string("value_states_163_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 value_states_163_groups_0 = const()[name = string("value_states_163_groups_0"), val = int32(1)];
            tensor<fp16, [256, 2048, 1, 1]> layers_27_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_27_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [256, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1619096640)))];
            tensor<fp16, [1, 256, 1, 8]> value_states_163_cast_fp16 = conv(dilations = value_states_163_dilations_0, groups = value_states_163_groups_0, pad = value_states_163_pad_0, pad_type = value_states_163_pad_type_0, strides = value_states_163_strides_0, weight = layers_27_self_attn_v_proj_weight_to_fp16, x = var_9597_cast_fp16_0)[name = string("value_states_163_cast_fp16")];
            tensor<int32, [4]> var_9649 = const()[name = string("op_9649"), val = tensor<int32, [4]>([1, 16, 128, 8])];
            tensor<fp16, [1, 16, 128, 8]> x_271_cast_fp16 = reshape(shape = var_9649, x = query_states_163_cast_fp16)[name = string("x_271_cast_fp16")];
            tensor<int32, [4]> var_9653 = const()[name = string("op_9653"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_9654_cast_fp16 = reshape(shape = var_9653, x = key_states_271_cast_fp16)[name = string("op_9654_cast_fp16")];
            tensor<int32, [4]> var_9660 = const()[name = string("op_9660"), val = tensor<int32, [4]>([1, 2, 128, 8])];
            tensor<fp16, [1, 2, 128, 8]> var_9661_cast_fp16 = reshape(shape = var_9660, x = value_states_163_cast_fp16)[name = string("op_9661_cast_fp16")];
            tensor<int32, [4]> value_states_165_perm_0 = const()[name = string("value_states_165_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 16, 128, 8]> var_9665_cast_fp16 = mul(x = x_271_cast_fp16, y = var_859_cast_fp16)[name = string("op_9665_cast_fp16")];
            tensor<int32, [2]> var_9666_split_sizes_0 = const()[name = string("op_9666_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_9666_axis_0 = const()[name = string("op_9666_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 16, 64, 8]> var_9666_cast_fp16_0, tensor<fp16, [1, 16, 64, 8]> var_9666_cast_fp16_1 = split(axis = var_9666_axis_0, split_sizes = var_9666_split_sizes_0, x = x_271_cast_fp16)[name = string("op_9666_cast_fp16")];
            fp16 const_331_promoted_to_fp16 = const()[name = string("const_331_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 8]> var_9668_cast_fp16 = mul(x = var_9666_cast_fp16_1, y = const_331_promoted_to_fp16)[name = string("op_9668_cast_fp16")];
            int32 var_9670 = const()[name = string("op_9670"), val = int32(-2)];
            bool var_9671_interleave_0 = const()[name = string("op_9671_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 8]> var_9671_cast_fp16 = concat(axis = var_9670, interleave = var_9671_interleave_0, values = (var_9668_cast_fp16, var_9666_cast_fp16_0))[name = string("op_9671_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> var_9672_cast_fp16 = mul(x = var_9671_cast_fp16, y = var_866_cast_fp16)[name = string("op_9672_cast_fp16")];
            tensor<fp16, [1, 16, 128, 8]> query_states_165_cast_fp16 = add(x = var_9665_cast_fp16, y = var_9672_cast_fp16)[name = string("query_states_165_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_9678_cast_fp16 = mul(x = var_9654_cast_fp16, y = var_859_cast_fp16)[name = string("op_9678_cast_fp16")];
            tensor<int32, [2]> var_9679_split_sizes_0 = const()[name = string("op_9679_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_9679_axis_0 = const()[name = string("op_9679_axis_0"), val = int32(-2)];
            tensor<fp16, [1, 2, 64, 8]> var_9679_cast_fp16_0, tensor<fp16, [1, 2, 64, 8]> var_9679_cast_fp16_1 = split(axis = var_9679_axis_0, split_sizes = var_9679_split_sizes_0, x = var_9654_cast_fp16)[name = string("op_9679_cast_fp16")];
            fp16 const_332_promoted_to_fp16 = const()[name = string("const_332_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_9681_cast_fp16 = mul(x = var_9679_cast_fp16_1, y = const_332_promoted_to_fp16)[name = string("op_9681_cast_fp16")];
            int32 var_9683 = const()[name = string("op_9683"), val = int32(-2)];
            bool var_9684_interleave_0 = const()[name = string("op_9684_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 8]> var_9684_cast_fp16 = concat(axis = var_9683, interleave = var_9684_interleave_0, values = (var_9681_cast_fp16, var_9679_cast_fp16_0))[name = string("op_9684_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> var_9685_cast_fp16 = mul(x = var_9684_cast_fp16, y = var_866_cast_fp16)[name = string("op_9685_cast_fp16")];
            tensor<fp16, [1, 2, 128, 8]> key_states_275_cast_fp16 = add(x = var_9678_cast_fp16, y = var_9685_cast_fp16)[name = string("key_states_275_cast_fp16")];
            tensor<int32, [4]> key_states_277_perm_0 = const()[name = string("key_states_277_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [1]> expand_dims_325 = const()[name = string("expand_dims_325"), val = tensor<int32, [1]>([27])];
            tensor<int32, [1]> expand_dims_326 = const()[name = string("expand_dims_326"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_328 = const()[name = string("expand_dims_328"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_329 = const()[name = string("expand_dims_329"), val = tensor<int32, [1]>([28])];
            int32 concat_219_axis_0 = const()[name = string("concat_219_axis_0"), val = int32(0)];
            bool concat_219_interleave_0 = const()[name = string("concat_219_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_219 = concat(axis = concat_219_axis_0, interleave = concat_219_interleave_0, values = (expand_dims_325, expand_dims_326, position_id, expand_dims_328))[name = string("concat_219")];
            tensor<int32, [1]> concat_220_values1_0 = const()[name = string("concat_220_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_220_values3_0 = const()[name = string("concat_220_values3_0"), val = tensor<int32, [1]>([0])];
            int32 concat_220_axis_0 = const()[name = string("concat_220_axis_0"), val = int32(0)];
            bool concat_220_interleave_0 = const()[name = string("concat_220_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_220 = concat(axis = concat_220_axis_0, interleave = concat_220_interleave_0, values = (expand_dims_329, concat_220_values1_0, cache_position_end, concat_220_values3_0))[name = string("concat_220")];
            tensor<int32, [4]> key_cache_internal_tensor_assign_28_stride_0 = const()[name = string("key_cache_internal_tensor_assign_28_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_28_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_28_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_28_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_28_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> key_cache_internal_tensor_assign_28_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_28_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> key_states_277_cast_fp16 = transpose(perm = key_states_277_perm_0, x = key_states_275_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [28, 2, 2048, 128]> key_cache_internal_tensor_assign_28_cast_fp16 = slice_update(begin = concat_219, begin_mask = key_cache_internal_tensor_assign_28_begin_mask_0, end = concat_220, end_mask = key_cache_internal_tensor_assign_28_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_28_squeeze_mask_0, stride = key_cache_internal_tensor_assign_28_stride_0, update = key_states_277_cast_fp16, x = coreml_update_state_108)[name = string("key_cache_internal_tensor_assign_28_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_28_cast_fp16, input = key_cache)[name = string("coreml_update_state_110_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_110 = read_state(input = key_cache)[name = string("coreml_update_state_110")];
            tensor<int32, [4]> value_cache_internal_tensor_assign_28_stride_0 = const()[name = string("value_cache_internal_tensor_assign_28_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_28_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_28_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_28_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_28_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
            tensor<bool, [4]> value_cache_internal_tensor_assign_28_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_28_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [1, 2, 8, 128]> value_states_165_cast_fp16 = transpose(perm = value_states_165_perm_0, x = var_9661_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [28, 2, 2048, 128]> value_cache_internal_tensor_assign_28_cast_fp16 = slice_update(begin = concat_219, begin_mask = value_cache_internal_tensor_assign_28_begin_mask_0, end = concat_220, end_mask = value_cache_internal_tensor_assign_28_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_28_squeeze_mask_0, stride = value_cache_internal_tensor_assign_28_stride_0, update = value_states_165_cast_fp16, x = coreml_update_state_109)[name = string("value_cache_internal_tensor_assign_28_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_28_cast_fp16, input = value_cache)[name = string("coreml_update_state_111_write_state")];
            tensor<fp16, [28, 2, 2048, 128]> coreml_update_state_111 = read_state(input = value_cache)[name = string("coreml_update_state_111")];
            tensor<int32, [4]> var_9755_begin_0 = const()[name = string("op_9755_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_9755_end_0 = const()[name = string("op_9755_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_9755_end_mask_0 = const()[name = string("op_9755_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_9755_cast_fp16 = slice_by_index(begin = var_9755_begin_0, end = var_9755_end_0, end_mask = var_9755_end_mask_0, x = coreml_update_state_110)[name = string("op_9755_cast_fp16")];
            tensor<int32, [2]> tile_54 = const()[name = string("tile_54"), val = tensor<int32, [2]>([1, 1])];
            int32 var_9758_axis_0 = const()[name = string("op_9758_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_9758_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_9758_cast_fp16_1 = split(axis = var_9758_axis_0, split_sizes = tile_54, x = var_9755_cast_fp16)[name = string("op_9758_cast_fp16")];
            tensor<int32, [4]> var_9765_begin_0 = const()[name = string("op_9765_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_9765_end_0 = const()[name = string("op_9765_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 128])];
            tensor<bool, [4]> var_9765_end_mask_0 = const()[name = string("op_9765_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 128]> var_9765_cast_fp16 = slice_by_index(begin = var_9765_begin_0, end = var_9765_end_0, end_mask = var_9765_end_mask_0, x = coreml_update_state_111)[name = string("op_9765_cast_fp16")];
            tensor<int32, [2]> tile_55 = const()[name = string("tile_55"), val = tensor<int32, [2]>([1, 1])];
            int32 var_9768_axis_0 = const()[name = string("op_9768_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1, 2048, 128]> var_9768_cast_fp16_0, tensor<fp16, [1, 1, 2048, 128]> var_9768_cast_fp16_1 = split(axis = var_9768_axis_0, split_sizes = tile_55, x = var_9765_cast_fp16)[name = string("op_9768_cast_fp16")];
            tensor<int32, [2]> var_9771_split_sizes_0 = const()[name = string("op_9771_split_sizes_0"), val = tensor<int32, [2]>([8, 8])];
            int32 var_9771_axis_0 = const()[name = string("op_9771_axis_0"), val = int32(1)];
            tensor<fp16, [1, 8, 128, 8]> var_9771_0, tensor<fp16, [1, 8, 128, 8]> var_9771_1 = split(axis = var_9771_axis_0, split_sizes = var_9771_split_sizes_0, x = query_states_165_cast_fp16)[name = string("op_9771")];
            bool attn_weights_433_transpose_x_0 = const()[name = string("attn_weights_433_transpose_x_0"), val = bool(false)];
            bool attn_weights_433_transpose_y_0 = const()[name = string("attn_weights_433_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_433_cast_fp16 = matmul(transpose_x = attn_weights_433_transpose_x_0, transpose_y = attn_weights_433_transpose_y_0, x = var_9758_cast_fp16_0, y = var_9771_0)[name = string("attn_weights_433_cast_fp16")];
            fp16 var_9774_to_fp16 = const()[name = string("op_9774_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_435_cast_fp16 = mul(x = attn_weights_433_cast_fp16, y = var_9774_to_fp16)[name = string("attn_weights_435_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_437_cast_fp16 = add(x = attn_weights_435_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_437_cast_fp16")];
            int32 var_9778 = const()[name = string("op_9778"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_439_cast_fp16 = softmax(axis = var_9778, x = attn_weights_437_cast_fp16)[name = string("attn_weights_439_cast_fp16")];
            bool var_9784_transpose_x_1 = const()[name = string("op_9784_transpose_x_1"), val = bool(true)];
            bool var_9784_transpose_y_1 = const()[name = string("op_9784_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> var_9784_cast_fp16 = matmul(transpose_x = var_9784_transpose_x_1, transpose_y = var_9784_transpose_y_1, x = attn_weights_439_cast_fp16, y = var_9768_cast_fp16_0)[name = string("op_9784_cast_fp16")];
            bool attn_weights_441_transpose_x_0 = const()[name = string("attn_weights_441_transpose_x_0"), val = bool(false)];
            bool attn_weights_441_transpose_y_0 = const()[name = string("attn_weights_441_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_441_cast_fp16 = matmul(transpose_x = attn_weights_441_transpose_x_0, transpose_y = attn_weights_441_transpose_y_0, x = var_9758_cast_fp16_1, y = var_9771_1)[name = string("attn_weights_441_cast_fp16")];
            fp16 var_9786_to_fp16 = const()[name = string("op_9786_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_443_cast_fp16 = mul(x = attn_weights_441_cast_fp16, y = var_9786_to_fp16)[name = string("attn_weights_443_cast_fp16")];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_445_cast_fp16 = add(x = attn_weights_443_cast_fp16, y = attn_mask_1_cast_fp16)[name = string("attn_weights_445_cast_fp16")];
            int32 var_9790 = const()[name = string("op_9790"), val = int32(-2)];
            tensor<fp16, [1, 8, 2048, 8]> attn_weights_cast_fp16 = softmax(axis = var_9790, x = attn_weights_445_cast_fp16)[name = string("attn_weights_cast_fp16")];
            bool attn_output_217_transpose_x_1 = const()[name = string("attn_output_217_transpose_x_1"), val = bool(true)];
            bool attn_output_217_transpose_y_1 = const()[name = string("attn_output_217_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 8, 8, 128]> attn_output_217_cast_fp16 = matmul(transpose_x = attn_output_217_transpose_x_1, transpose_y = attn_output_217_transpose_y_1, x = attn_weights_cast_fp16, y = var_9768_cast_fp16_1)[name = string("attn_output_217_cast_fp16")];
            int32 var_9798 = const()[name = string("op_9798"), val = int32(1)];
            bool attn_output_219_interleave_0 = const()[name = string("attn_output_219_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 8, 128]> attn_output_219_cast_fp16 = concat(axis = var_9798, interleave = attn_output_219_interleave_0, values = (var_9784_cast_fp16, attn_output_217_cast_fp16))[name = string("attn_output_219_cast_fp16")];
            tensor<int32, [4]> var_9802_perm_0 = const()[name = string("op_9802_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_9807 = const()[name = string("op_9807"), val = tensor<int32, [4]>([1, 2048, 1, 8])];
            tensor<fp16, [1, 16, 128, 8]> var_9802_cast_fp16 = transpose(perm = var_9802_perm_0, x = attn_output_219_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 2048, 1, 8]> attn_output_cast_fp16 = reshape(shape = var_9807, x = var_9802_cast_fp16)[name = string("attn_output_cast_fp16")];
            string hidden_states_273_pad_type_0 = const()[name = string("hidden_states_273_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_273_strides_0 = const()[name = string("hidden_states_273_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_273_pad_0 = const()[name = string("hidden_states_273_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_273_dilations_0 = const()[name = string("hidden_states_273_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_273_groups_0 = const()[name = string("hidden_states_273_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 2048, 1, 1]> layers_27_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_27_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1620145280)))];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_273_cast_fp16 = conv(dilations = hidden_states_273_dilations_0, groups = hidden_states_273_groups_0, pad = hidden_states_273_pad_0, pad_type = hidden_states_273_pad_type_0, strides = hidden_states_273_strides_0, weight = layers_27_self_attn_o_proj_weight_to_fp16, x = attn_output_cast_fp16)[name = string("hidden_states_273_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_275_cast_fp16 = add(x = hidden_states_269_cast_fp16, y = hidden_states_273_cast_fp16)[name = string("hidden_states_275_cast_fp16")];
            int32 var_9833 = const()[name = string("op_9833"), val = int32(1)];
            fp16 const_337_promoted_to_fp16 = const()[name = string("const_337_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_9835_cast_fp16 = mul(x = hidden_states_275_cast_fp16, y = const_337_promoted_to_fp16)[name = string("op_9835_cast_fp16")];
            bool doubled_221_interleave_0 = const()[name = string("doubled_221_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_221_cast_fp16 = concat(axis = var_9833, interleave = doubled_221_interleave_0, values = (hidden_states_275_cast_fp16, var_9835_cast_fp16))[name = string("doubled_221_cast_fp16")];
            tensor<int32, [1]> out_111_axes_0 = const()[name = string("out_111_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4096]> out_111_gamma_0_to_fp16 = const()[name = string("out_111_gamma_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1628533952)))];
            fp16 var_9845_to_fp16 = const()[name = string("op_9845_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_111_cast_fp16 = layer_norm(axes = out_111_axes_0, epsilon = var_9845_to_fp16, gamma = out_111_gamma_0_to_fp16, x = doubled_221_cast_fp16)[name = string("out_111_cast_fp16")];
            tensor<int32, [2]> var_9856_split_sizes_0 = const()[name = string("op_9856_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_9856_axis_0 = const()[name = string("op_9856_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> var_9856_cast_fp16_0, tensor<fp16, [1, 2048, 1, 8]> var_9856_cast_fp16_1 = split(axis = var_9856_axis_0, split_sizes = var_9856_split_sizes_0, x = out_111_cast_fp16)[name = string("op_9856_cast_fp16")];
            string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])];
            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, [6144, 2048, 1, 1]> layers_27_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_27_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1628542208)))];
            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_27_mlp_gate_proj_weight_to_fp16, x = var_9856_cast_fp16_0)[name = string("input_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> var_9873_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_9873_cast_fp16")];
            string var_9879_pad_type_0 = const()[name = string("op_9879_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_9879_strides_0 = const()[name = string("op_9879_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9879_pad_0 = const()[name = string("op_9879_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9879_dilations_0 = const()[name = string("op_9879_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_9879_groups_0 = const()[name = string("op_9879_groups_0"), val = int32(1)];
            tensor<fp16, [6144, 2048, 1, 1]> layers_27_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_27_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [6144, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1653708096)))];
            tensor<fp16, [1, 6144, 1, 8]> var_9879_cast_fp16 = conv(dilations = var_9879_dilations_0, groups = var_9879_groups_0, pad = var_9879_pad_0, pad_type = var_9879_pad_type_0, strides = var_9879_strides_0, weight = layers_27_mlp_up_proj_weight_to_fp16, x = var_9856_cast_fp16_0)[name = string("op_9879_cast_fp16")];
            tensor<fp16, [1, 6144, 1, 8]> x_cast_fp16 = mul(x = var_9873_cast_fp16, y = var_9879_cast_fp16)[name = string("x_cast_fp16")];
            string hidden_states_277_pad_type_0 = const()[name = string("hidden_states_277_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> hidden_states_277_strides_0 = const()[name = string("hidden_states_277_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> hidden_states_277_pad_0 = const()[name = string("hidden_states_277_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> hidden_states_277_dilations_0 = const()[name = string("hidden_states_277_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 hidden_states_277_groups_0 = const()[name = string("hidden_states_277_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 6144, 1, 1]> layers_27_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_27_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [2048, 6144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1678873984)))];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_277_cast_fp16 = conv(dilations = hidden_states_277_dilations_0, groups = hidden_states_277_groups_0, pad = hidden_states_277_pad_0, pad_type = hidden_states_277_pad_type_0, strides = hidden_states_277_strides_0, weight = layers_27_mlp_down_proj_weight_to_fp16, x = x_cast_fp16)[name = string("hidden_states_277_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states_cast_fp16 = add(x = hidden_states_275_cast_fp16, y = hidden_states_277_cast_fp16)[name = string("hidden_states_cast_fp16")];
            int32 var_9895 = const()[name = string("op_9895"), val = int32(1)];
            fp16 const_339_promoted_to_fp16 = const()[name = string("const_339_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2048, 1, 8]> var_9897_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_339_promoted_to_fp16)[name = string("op_9897_cast_fp16")];
            bool doubled_225_interleave_0 = const()[name = string("doubled_225_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4096, 1, 8]> doubled_225_cast_fp16 = concat(axis = var_9895, interleave = doubled_225_interleave_0, values = (hidden_states_cast_fp16, var_9897_cast_fp16))[name = string("doubled_225_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(1704039872)))];
            fp16 var_9907_to_fp16 = const()[name = string("op_9907_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4096, 1, 8]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_9907_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_225_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [2]> var_9918_split_sizes_0 = const()[name = string("op_9918_split_sizes_0"), val = tensor<int32, [2]>([2048, 2048])];
            int32 var_9918_axis_0 = const()[name = string("op_9918_axis_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 8]> hidden_states, tensor<fp16, [1, 2048, 1, 8]> var_9918_cast_fp16_1 = split(axis = var_9918_axis_0, split_sizes = var_9918_split_sizes_0, x = out_cast_fp16)[name = string("op_9918_cast_fp16")];
        } -> (hidden_states);
}